New Developments in Sensing Technology for Structural Health Monitoring

Lecture Notes in Electrical Engineering Volume 96 Subhas Chandra Mukhopadhyay New Developments in Sensing Technology...

Author: Subhas C Mukhopadhyay

339 downloads 1380 Views 14MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

Lecture Notes in Electrical Engineering Volume 96

Subhas Chandra Mukhopadhyay

New Developments in Sensing Technology for Structural Health Monitoring

ABC

Prof. Subhas Chandra Mukhopadhyay Massey University 12 Woodgate Court Palmerston North New Zealand E-mail: [email protected]

ISBN 978-3-642-21098-3

e-ISBN 978-3-642-21099-0

DOI 10.1007/978-3-642-21099-0 Lecture Notes in Electrical Engineering

ISSN 1876-1100

Library of Congress Control Number: 2011928067 c 2011 Springer-Verlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typeset & Coverdesign: Scientific Publishing Services Pvt. Ltd., Chennai, India. Printed on acid-free paper 987654321 springer.com

Guest Editorial

In recent times several incidents of bridge/buildings collapse took place in different parts of the world. After these accidents it has become paramount importance of early detection of the health of structures and the sensors must have intelligent features to detect the problem. It is expected that the special issue have provided many new ideas of detection and inspection of the health of structures which are very important for human being and society. There is an urgent need to design, develop, fabricate of different types of sensors and sensing technology based on noninvasive techniques to determine the integrity of a material, component or structure or quantitatively measure some characteristics of the systems to prevent catastrophic failure. So in short the fabricated sensor systems should be able to inspect or measure without doing any harm or damage of the system. Not only the monitoring of structural health the applications of the developed sensing systems are necessary at almost any stage in the production or the life cycle of a component in many years such as civil engineering, metal industry, transportation, power stations, inspection of pipes and piping systems in industrial plants, fatigue estimation in aircraft surface and other parts and in many other areas. Many different sensing techniques available with different characteristics are available for these inspection areas. The following are the most commonly used: Magnetic, Ultrasonic, Acoustic, Radiography, Eddy current and X-ray. The sensors to be used entirely depend entirely on the specific application. The proposed Special Issue has focussed on the different aspects of sensing technology, i.e. high reliability, adaptability, recalibration, information processing, data fusion, validation and integration of novel and high performance sensors specifically aims to use to inspect mechanical health of structure and similar applications. The book, on one hand, illustrates theoretical aspects and applications, and it displays new criteria in characterizing raw data of SHM, at the other hand. Characterization is a key issue since it allows to know the performances of devices and systems described in the book by showing some statistics and result representation. The book contains 15 contributions from experts working on the topic and under different approaches and aspects; these co-ordinated approaches are the true richness of the book. The editor gracefully thanks the contributors for contribution included in this special issue. The editor hopes this special issue will be a very useful for readers with experience who can breathe fresh life into their research. Subhas Chandra Mukhopadhyay, Guest Editor School of Engineering and Advanced Technology (SEAT), Massey University (Turitea Campus) Palmerston North, New Zealand [email protected]

VI

Guest Editorial

Dr. Subhas Chandra Mukhopadhyay graduated from the Department of Electrical Engineering, Jadavpur University, Calcutta, India in 1987 with a Gold medal and received the Master of Electrical Engineering degree from Indian Institute of Science, Bangalore, India in 1989. He obtained the PhD (Eng.) degree from Jadavpur University, India in 1994 and Doctor of Engineering degree from Kanazawa University, Japan in 2000. During 1989-90 he worked almost 2 years in the research and development department of Crompton Greaves Ltd., India. In 1990 he joined as a Lecturer in the Electrical Engineering department, Jadavpur University, India and was promoted to Senior Lecturer of the same department in 1995. Obtaining Monbusho fellowship he went to Japan in 1995. He worked with Kanazawa University, Japan as researcher and Assistant professor till September 2000. In September 2000 he joined as Senior Lecturer in the Institute of Information Sciences and Technology, Massey University, New Zealand where he is working currently as an Associate professor. His fields of interest include Sensors and Sensing Technology, Electromagnetics, control, electrical machines and numerical field calculation etc. He has authored over 200 papers in different international journals and conferences, edited nine conference proceedings. He has also edited eight special issues of international journals as guest editor and ten books with Springer-Verlag. He is a Fellow of IEEE (USA), a Fellow of IET (UK), an associate editor of IEEE Sensors journal and IEEE Transactions on Instrumentation and Measurements. He is in the editorial board of e-Journal on Non-Destructive Testing, Sensors and Transducers, Transactions on Systems, Signals and Devices (TSSD), Journal on the Patents on Electrical Engineering, Journal of Sensors. He is the coEditor-in-chief of the International Journal on Smart Sensing and Intelligent Systems (www.s2is.org). He is in the technical programme committee of IEEE Sensors conference, IEEE IMTC conference and IEEE DELTA conference and numerous other conferences. He was the Technical Programme Chair of ICARA 2004, ICARA 2006 and ICARA 2009. He was the General chair/co-chair of ICST 2005, ICST 2007, IEEE ROSE 2007, IEEE EPSA 2008, ICST 2008, IEEE Sensors 2008, ICST 2010 and IEEE Sensors 2010. He has organized the IEEE Sensors conference 2009 at Christchurch, New Zealand during October 25 to 28, 2009 as General Chair. He is the Chair of the IEEE Instrumentation and Measurement Society New Zealand Chapter. He is a Distinguished Lecturer of the IEEE Sensors Council.

Contents

Sensors and Technologies for Structural Health Monitoring: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . S.C. Mukhopadhyay, I. Ihara Self-sustaining Wireless Acoustic Emission Sensor System for Bridge Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ´ Akos Lédeczi, Péter V¨ olgyesi, Eric Barth, Andr´ as N´ adas, Alexander Pedchenko, Thomas Hay, Subash Jayaraman

1

15

Deformation Detection in Structural Health Monitoring . . . . . Pierantonio Merlino, Antonio Abramo

41

MEMS Strain Sensors for Intelligent Structural Systems . . . . . Debbie G. Senesky, Babak Jamshidi

63

A Pattern-Based Framework for Developing Wireless Monitoring Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . James Brusey, Elena Gaura, Roger Hazelden

75

Distributed Brillouin Sensor Application to Structural Failure Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F. Ravet

93

Sensing Network Paradigms for Structural Health Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 C.R. Farrar, G. Park, M.D. Todd Reflectometry for Structural Health Monitoring . . . . . . . . . . . . . 159 Cynthia Furse Sensor Fusion in Transportation Infrastructure Systems Using Belief Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 Stephen Mensah, Nii O. Attoh-Okine, Ardeshir Faghri Pulsed Eddy Current Thermography and Applications . . . . . . . 205 G.Y. Tian, J. Wilson, L. Cheng, D.P. Almond, E. Kostson, B. Weekes

VIII

Contents

The Use of Optical Fibre Sensors in Dam Monitoring . . . . . . . . 233 Ian Platt, Michael Hagedorn, Ian Woodhead Optical Sensors Based on Fiber Bragg Gratings for Structural Health Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 P. Antunes, H. Lima, N. Alberto, L. Bilro, P. Pinto, A. Costa, H. Rodrigues, J.L. Pinto, R. Nogueira, H. Varum, P.S. André Polymer Optical Fiber Sensors in Structural Health Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297 Sascha Liehr Optical Fiber Sensors for Structural Health Monitoring . . . . . 335 Alayn Loayssa Sensors Systems, Especially Fibre Optic Sensors in Structural Monitoring Applications in Concrete: An Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 S.K.T. Grattan, S.E. Taylor, P.M.A. Basheer, T. Sun, K.T.V. Grattan Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427

Sensors and Technologies for Structural Health Monitoring: A Review S.C. Mukhopadhyay1 and I. Ihara2 1

School of Engineering and Advanced Technology Massey University, Palmerston North, New Zealand [email protected] 2 Department of Mechanical Engineering Nagaoka University of Technology, Nagaoka, Japan [email protected]

Abstract. Incidents such as building and bridge collapse are on rise in many parts of the world without little apparent warning. Due to the increase number of incidents it has become of increasingly paramount importance to develop methods detecting the degradation or damage that result in these events. Thus, buildings and critical infrastructure could be monitored, much like a patient in a hospital, for signs of degradation or impending disability or collapse. The sensors are very important to know the state of the health of the structures and technologies are like human brains to analyze the abnormal situation. This chapter will provide a review of different available sensors and technologies to be used for monitoring the health of structures.

1 Introduction and Literature Review Intelligent sensors and technologies that are able to take a potentially diverse array of data and create a picture of the structure’s condition will help to determine the early detection of damage from natural hazards or other events. Thus, the sensors must have access to or contain intelligent features to detect the problem. It is therefore important to know wide varieties of sensors and technologies for Structural Health Monitoring (SHM) which can be deployed for the detection and inspection of structures to increase their safety and reliability. The reported sensor and technologies should be able to inspect or measure without doing any harm or damage of the structure. They should also be robust to poor signal-to-noise ratio compared to the level of damage they are trying to detect in these critical infrastructures. Finally, they need to be highly reliable and operate without input for long periods of time, potentially over years. A lot of research articles have been reported on monitoring health of structures. A structural health monitoring system based on wireless sensor nodes equipped with inexpensive strain gauges has been proposed [1]. Due to the deployment of multi-hop technique the performance of the system is not limited. Strain gauges are very popular in SHM as they are inexpensive, easy to install and having good sensitivity to detect potential danger or collapse of a building or structure. The developed system has been tested with simulated structure. S.C. Mukhopadhyay (Ed.): New Developments in Sensing Technology for SHM, LNEE 96, pp. 1–14. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

2

S.C. Mukhopadhyay and I. Ihara

MEMS inertial sensors [2] including an acceleration sensor and an angular velocity sensor (gyroscope) can be used as a popular device for monitoring the health of structure due to their miniaturized size, low cost, mass production and three-dimensional detection. An impedance measurement system for lead zirconate titanate (PZT) ceramics based SHM has been reported in [3]. The PZT sensors are inexpensive, small, light weight, require low power, less sensitive to temperature variation and provide a linear response under low electric field. The importance of monitoring health of aerospace structure using optical sensors was considered more than a decade back as was reported by Foote and Read [4]. It states that with the help of a smart sensor network, the stress and strains induced in the aircraft and possible degradation occurred since last inspection can be known clearly. Fibre optic accelerometer based monitoring of civil engineering infrastructure and damage detection of concrete slab has been reported by Kim and Feng [5]. The sensor system integrates Moire fringe phenomenon with fibre optics to achieve accurate and reliable measurement. Fibre optic sensors emerged as an important technology to evaluate structural integrity [6]. The strain along the fibre length provides distributed information about mechanical state of the structure. Bo-lin et. al., [7] have described some works and applications of new sensors such as optical fibre sensors, piezoelectric sensors, MEMS sensors, wireless sensing system etc. for aircraft structural health monitoring. The experimental works have been carried out in laboratory conditions and some more works are required to integrate the sensors to the structures effectively, determination of optimum number of sensors and their location and enhancement of the reliability of the sensors in order to survive the rugged environments. In [8] a structural health monitoring system using wireless sensor network consisting of 17 sensor nodes, a base station and a processing computer has been implemented. The acceleration data synchronously sampled from each sensor node are transported to a data processing computer through a base station. A time division multiple access (TDMA) approach has been proposed to reduce the packet collision and energy consumption. The experimental works on the design and implementation of an innovative technological framework for monitoring critical structures in Italy has been reported [9]. The use of wireless sensors networks allowed for a pervasive observation over the sites of interest to minimize the potential damages that natural phenomenon may cause to architectural or engineering works. The temperature, relative humidity, linear strain and 3-axis acceleration sensors are used for the measurement of observed parameters. A SHM flexible testbed system has been developed for detecting high-velocity impacts in the skin of a structure [10]. The system is a large sensor network containing about two hundred nodes, each of which contains multiple sensors. The testbed is used for studying wide range of SHM applications. The configurations of a novel wireless system for infrastructure health monitoring has been proposed and developed with a special attention to the low frequency characteristics of the wireless transmission [11].

Sensors and Technologies for Structural Health Monitoring: A Review

3

Sensors deployed for monitoring bridges, buildings etc. always face a constraint from energy consideration. A novel wireless sensor system has been presented in [12] that harvests vibrations of the bridge created by passing traffic, which is converted into usable energy by means of a linear electromagnetic generator. In the particular design [12], harvesting of power up to 12.5 mW in the resonant mode with an excitation frequency of 3.1 Hz has been reported. A field study of monitoring the ambient vibration using 60 accelerometers interfaced with 30 wireless sensor nodes operating within one or two simultaneously star topology network has been reported [13]. It is envisioned that the reported system can address short-term and long-term management and condition assessment needs for highway bridges. In [14], a novel sensor network architecture for SHM has been presented. The system is based on contactless sensors that make use of near-field coupling to both sense the structure displacement and deploy a local communication network. A simple custom-built gages based detection of cracks in critical structural elements and its design, implementation and experimental evaluation of a WSN for real-time SHM has been reported [15]. The paper [15] has shown that a variety of low-cost, off-the-shelf data acquisition/communication devices can be used to support remote monitoring by a control centre. The assessment of the developed system done for a full-scale three-story reinforced concrete building that was tested under lateral forces emulating forces induced by earthquakes. P.F.dC. Antunes et. al., [16] have reported the implementation of an optical accelerometer unit based on fiber Bragg gratings, suitable to monitor structures with frequencies up to 45 Hz. The developed system has been used to estimate the eigenfrequencies of a steel foot bridge structure of total length of 300 m. Bragg grating-based optical fiber sensors integrated into carbon fiber polymer reinforcement (CFPR) rod have been used to measure strains in concrete structures [17]. It has been concluded from experimental results that the effective strain measurement can be obtained from the different sensors mounted along the rod. From the results it can be concluded that in-situ monitoring of strains in different engineering structure is possible. In [18] comparative test results between the performance of electrical resistance strain gauges (ERSG) and fiber-optic sensors (FOS) based on in-fiber Bragg grating technology for monitoring health of structures are reported. The results have shown a close comparison of the data obtained between different methods of strain measurement. Micro-Opto-Electro-Mechanical Systems (MOEMS) acoustic sensors have been employed to detect acoustic emissions (AE) for Structural Health Monitoring (SHM) [19]. Acoustic sensing cantilevers (~ 200 x 100 x 50 μm) with variable frequency response, directionality and dynamic range have been fabricated in large quantity using a novel non-silicon process. The packaged sensors are low-cost, easy-toinstall and ElectroMagnetic Interference (EMI) free during operation. The acoustic sensors’ broadband sensitivity is demonstrated by standard structural break tests. A wireless embedded system that performs active ultrasonic SHM has been reported in [20]. The proposed Shimmer platform is an autonomous, battery-less

4


system, powered by an energy harvester, which uses 16 piezoelectric sensors/actuators attached to the structure surface for SHM analysis. The collected data (4MB) are analyzed by the on-board DSP and the result is transmitted using the 6LoWPAN protocol with IEEE802.15.4 transceiver working at 915MHz. The significant challenge is running such intense analysis on the collected data with only energy harvesting as a power source. Sensors used for monitoring the health of structures are critical and their damage will create important issues in SHM. Y. Matsushiba and H. Nishi have proposed data-distributed fault-tolerant sensor network system for a SHM system [21-22]. The proposed system consists of three functions, PING (Packet InterNet Groper) based Network Monitoring function (PNM), backup node selecting function (BNS) and socket communication function. The proposed system has been implemented on PCs and has been practically evaluated in a laboratory environment on test bed system. The inherent limitations of WSN such as low-bandwidth wireless communication, limited resources on wireless sensors nodes need to be addressed for a successful SHM system. A multi-scale strategy in WSNs for SHM has been proposed in [23]. The approach, called the Auto-Correlation Function and Cross-Correlation Function (ACFCCF), utilizes the autocorrelation function of individual sensor node to detect the existence of damage and the cross-correlation function of designated node pairs to obtain damage location. The signal processing issues related to SHM have been described in [24]. The key components of the SHM process include data acquisition and normalization, feature extraction and information condensation and statistical model development. Piezoelectric based distributed sensors are embedded to investigate the deformation and deflection of the buried pipes due to unexpected and external loadings [25]. The vibration and frequency response of a modeled pipeline integrated with piezoelectric sensors has been investigated to identify, locate and quantify the structural performance of the system. The requirements of sensors for monitoring the health of structures are to be cheap, replaceable, durable, low-power requirements and on-site artificial intelligence which will be useful to distinguish the abnormal behavior from normal one. The damage detection techniques should be developed which can recognize when damage has occurred and provide direction to the location of the damage [26]. The SHiMmer, a wireless platform for sensing and actuation that combines localized processing with energy harvesting to provide long-lived structural health monitoring has been reported in [27]. It has been reported that with the use of super-capacitor the life-cycle of the node has been significantly extended. The development of an integrated structural health monitoring and reporting (SHMR) system for use on Navy aircraft has been discussed in [28]. Wireless sensors included strain gauges, accelerometers and thermocouples and wired sensors included gyroscopes, accelerometers and magnetometers have been used. The data from an embedded Global Positioning System (GPS) provided position, velocity and precise timing information.


5

The authors have reported the development of an all-digital excitation and sensing techniques to reduce the size of the hardware and power requirements of the system used for SHM [29]. For an energy harvesting WSN, it is important for the task scheduler to adapt the task complexity and maximize the accuracy of the tasks within the constraints of limited energy reserves. A task scheduler has been proposed for the sensing systems for SHM [30]. It is based on a Linear Regression Model embedded with Dynamic Voltage and Frequency Scaling (DVFS) functionality. The feasibility of using electrical reflectometry method for fault location on concrete anchors has been explored in [31]. Concrete dams and other large civil structures utilize steel cable anchors to improve strength and stability. Reflectometry methods are suitable to determine the location of faults and quantization of possible deterioration on concrete anchors. To employ WSN for SHM application, system requirements are need to be considered. In [32] the system requirements posed by SHM applications have been considered to assess potential candidates for the protocols in WSN for SHM. The authors have concluded that none of the commonly available protocols satisfy all the requirements associated with SHM systems. So there is a necessity to modify the existing protocols or it may be good to design an entirely new protocol to completely satisfy the requirements of WSN for SHM applications. T. Harms, S. Sedigh and F. Bastianini [33] have described a complete overview of emerging wireless sensor networks for autonomous SHM systems, their application, the power use and sources needed to support autonomy and the type of communication that allows remote monitoring. A proper analysis of raw data from sensors is very important to conclude the real and accurate prediction of health of structures. Many algorithms are proposed for the analysis. Ling Yu et.al., [34, 35] have proposed a Principal Component Analysis (PCA) based and an Ant Colony Optimization (ACO) based algorithm techniques to apply in SHM systems. The authors have summarized the most important measurement results of standard POF (Polymer Optical Fiber) and PF GI POF (PerFluorinated GradedIndex Polymer Optical Fiber), their strain and external disturbances with respect to their applicability of SHM [36]. The authors have presented an algorithm [37] for real-time SHM during earthquake events using only acceleration measurements and infrequently measured displacement motivated by global positioning system. The developed algorithm has identified a nonlinear baseline model including hysteretic dynamics and permanent deformation using convex integral-based fitting methods and piecewise linear least squares fitting. The authors [38] have presented a prototype wireless system for the detection of active fatigue cracks in aging railway bridges in real-time. The system is based on a small low-cost sensor node, called an AEPod, that has four acoustic emission (AE) channels and a strain channel for sensing, as well as the capability to communicate in a wireless fashion with other nodes and a base station. A new type of passive wireless sensor based on resonant RF cavities has been reported [39]. The significant problems in the installation and long term use of

6


wired sensors have been eliminated. The resonant frequency is being modulated by a measurand. A probe inside the cavity couples RF signals from the cavity to an externally attached antenna. The sensor can then be interrogated remotely using microwave pulse-echo techniques. A wireless, multisensory inspection system for nondestructive evaluation of materials has been described [40].

2 Characteristics of Sensors for Monitoring Health of Structures The sensors are the fundamental element for monitoring the health of structures. There are different types of sensors available and to be chosen depending on the applications. A few sensors such as strain gauges, accelerometers, temperature, acoustic emission sensors are very commonly used sensors used for monitoring the health of structures. Nowadays, the fibre optic based sensor systems are becoming very popular due to their different advantages compared to other sensors. The selection of sensors, cost, number of sensors and their placements, protection against mechanical and chemical damage, reduction of noise, and the collection of more representative data are the few things considered for the sensors used for health monitoring. The sensitivity of sensors to moisture and humidity is another concern, especially when long-term measurement is planned, particularly in a harsh environment. Special provisions are often needed to protect the sensors in order to obtain acceptable measurements. Since the sensors are planned to be used for a long duration, the energy harvesting may also need to be considered. The sensors considered for health monitoring of structures are usually smart wireless sensors as wired sensors may not be a cheap and simple option for this type of applications. The sensors along with signal conditioning in combination with a microcontroller/microprocessor all come in the same package and can be defined as smart sensors or smart wireless sensors. Usually the smart sensors have the ability to compensate for random errors, can adapt to changes in the environment, can adjust non-linearities to give a linear output, have the provision of self-calibration and self-diagnosis of fault. The smart sensors have their own standard, IEEE 1451 so that they can be used in a ‘plug-and-play’ manner. The following characteristics are very important for the selection of sensors used for monitoring the health of structures. a) Range It is defined as the limits between which the inputs of the sensor can vary. It is very important for the sensors used for health monitoring of structures as the maximum input applied to the sensor may be unknown in many instances. The sensors should not be damaged at that abnormal condition. b) Sensitivity The sensors used for health monitoring should be sensitive enough to give correct information on the effect of the input signal. The sensitivity is the relationship


7

between output and input and is also used to indicate about the change of output to inputs other than being measured, such as environmental parameter changes. It is desirable that the sensitivity of the sensors to the environmental parameter changes is ideally zero or should be very small so that it can be easily negligible. c)

Accuracy

It is a measure of the closeness of the actual output to the ideal output of the sensor. It is an indication of the extent by which the measurement is wrong. It is the summation of all the possible errors that are likely to occur and it also depends on the calibration method. The accuracy may be represented either in absolute value or may be in percentage of the full range output. d) Stability The sensors used for SHM are in service continuously over many years. The sensors should be stable enough to give the same output for a constant input over a period of time. With respect to stability, a term ‘drift’ is used to describe the change in output that occurs over time. It is expressed as a percentage of the full range output. e)

Repeatability

This is very important for any sensors especially for the sensors used for SHM. It is the ability to give the same amount of output for repeated applications to the same amount of input. It is also termed as ‘reproducibility’. The error is usually expressed as a percentage of the full range output:

Repeatability=

Maximum Output (for an input) - Minimum Output (for the same input) *100 Full Range

It is expected that the repeatability of the sensors used for SHM should be better than 0.01%. f) Static and dynamic characteristics While the sensors are used for SHM, the static and dynamic characteristics of the sensors such as rise time, time constant and settling time should be looked into for selection. In some situation the slow response of the sensors may not be very critical for SHM applications. While the sensors are subjected to a dynamic input condition, the response should be free from hysteresis. g) Energy Harvesting The sensors employed for SHM are used for many years. So it may be a good idea to investigate some kind of energy harvesting option so that the sensors will be self-sufficient.

8


h) Compensation due to change of temperature and other environmental parameters The responses of the sensors are usually affected due to change of ambient temperature, humidity and other environmental parameters. To reduce the effect of the external influences, adequate compensation schemes must be included in the signal conditioning part of the sensors.

3 Analyzing Techniques for SHM The sensors raw data are analyzed to derive the information of the status of the health of the structures. Usually the sensors are wireless sensors distributed over a large area. The raw data can be collected from the sensors through wireless communication and all data can be gathered in the central processor situated at a far distant. For the transmission of data, a comparison has been made between a single-hop data transmission and multi-hop data transmission [1]. It has been shown that the throughput of a single source node decreases as the number of hops increase. This is due to the interference from 2- or 3-hop nodes, even though the radio range is only one hop [1]. One of the requirements of employing WSN for SHM is that the amount of wireless communication required by the algorithm to be as minimum as possible to save energy and decrease packet loss rate [22]. The sensor node should be light weight. Moreover, the resource consuming algorithms are not suitable to be implemented in smart sensor nodes. The whole software design should be such that is should be able to accommodate energy saving strategies like wakeup and sleep scheduling. The algorithm should be developed to minimize false-positive and false-negative indication of damage. The software should be able to detect damage at an early stage and should also provide the location of the damage. A combination of ACF (Auto-correlation function) and CCF (Cross-correlation function) can be useful for SHM. If there is any damage in the structure, the ACF of the obtained time series will be different from those obtained from the undamaged structure. It is expected that the ACF is sensitive to damage but not to input/environmental changes. If the ACF algorithm provides indication of damage then the CCF is used to locate the damage. The sensor nodes deployed on the structure are divided as node pairs. Each node pair covers an area of the structure where the node pair is located. If damage occurs in that area, the dynamic relationship of the two nodes in the node pair will be changed. The CCF of the two nodes will be altered correspondingly. To obtain correct information from optical fibre based strain sensors a considerable effort in signal de-multiplexing is required to realize practical systems [4]. The strain measurements were made separately recording the wavelength of the maxima in reflection from each Bragg grating. Two forms of electronically tunable optical filters were developed as de-multiplexing systems. One method which is very useful in SHM analysis is distributed processing – network intelligence [10]. The intelligent sensing network must produce intelligent responses to the sensed environment and it should be on-line. The


9

response should be produced by self-organization: it is emergent behavior of the system. The SHM problem involves multiple inter-related hierarchical sub-tasks (e.g. damage detection, evaluation, diagnosis, prognosis and repair). The approach assumes that single cells may make fast and automatic responses to critical emergencies, while collection of cells may solve more complex hierarchical tasks including [10]: i)

Self-calibration and discrimination between component and sensor failures; ii) Formation of a dynamic artificial neural network, characterizing the nature of possible damage and producing a self-organizing diagnosis; iii) Self-scheduling of secondary inspections, maintenance or corrective actions based on information from the network while issuing warnings; iv) Direction of recovery resources, human or robotic, to the repair site [10].

Determining desirable and quantitative information from the raw data observed by SHM may be equivalent to solving an inverse problem. Optimization procedures are often used to solve such inverse problem because optimum values are easier to obtain the exact values, and good enough for practical use. Ant Colony Optimization (ACO) algorithms has been proposed to be sued for SHM analysis [10, 35, 36]. The algorithms use the ability of agents to interact indirectly through changes in their environment by depositing pheromones and forming a pheromone trail. A form of autocatalytic behavior – allelomimesis: the probability with which an ant chooses a trail increases the number of ants that choose the same path in the past has also been employed. In the algorithm, ants are implemented as communication packets, policies are implemented via appropriate message passing, cells are responsible for interpreting received packets or sending packets. Since ants cannot move into the cells with broken communication links, they are supposed to find the shortest paths around them using positively reinforced pheromone trails. In general, the ACO approach attempts to solve an optimization problem by iterating the following two steps: i)

Candidate solutions are constructed in a probabilistic way by using a probability distribution over the search space, ii) the candidate solutions are used to modify the probability distribution in a way that is deemed to bias future sampling towards high quality solutions. An adaptive sensing approach based monitoring can be effective method of SHM [13]. The strain sensors can be used as the primary asset for health indication through schedules-based experimental load ratings to assess structural deterioration and determine up-to-date structural capacity. During the intermediate periods between load ratings, sensor can monitor continuously to detect any anomalies to indicate any external failure triggering events. The approach is well suited to wireless network based instrumentation as the system demands are within the power and bandwidth limitations. During the majority of time, vibration would be monitored locally by the sensor nodes to maintain ultralow- power consumption.

10


For successful SHM, fault tolerant sensor network for SHM has to be developed [21, 22, 24]. To implement this, distributed and data shared sensor network is indispensable. It is difficult to implement complex data sharing method in sensor network system due to its limited processing power and network bandwidth. The simple and effective data sharing method for fault-tolerance is inter-sharing of sampled data among sensor nodes. This sharing method may be influenced by the topology of the sensor network system which can be changed dynamically by both faulty nodes and link disconnections. The sensor node sends data to server at a regular or specific time. If the data backup timing is synchronized with the time when sensor node sends data to server, network throughput and processing load of sensor node will be increased rapidly according to the growth of data sharing throughput. Backup or data sharing node can be selected from all nodes of the sensor network. This means that the latency and throughput of backup is changeable by the distance between original node and backup node. If backup node is close to original node, synchronized backup is preferable. Thus, backup rate or timing should be controlled by the distance. It is not easy to measure damage directly from the sensors’ data obtained in SHM [23]. Many parameters must be measured throughout the structure and should be utilized to assess the health of the structure. It is important to know the type, number and placement of sensors on the structure, which are problem dependent. The following methods of analysis are quite common in relation to SHM: a)

Normalization: It is applied to account for changes in structural response due to environmental conditions or structural loads which are not associated with any structural damage. Some parameters, such as elastic modules of a structure, being temperature- dependent may have a significant effect on the dynamic of the structure. Normalization is required to ensure that the changes in the dynamics of the structure are due to change in temperature or any other environmental parameters, and is not interpreted as damage. Gain normalization is utilized by dividing the response by its peak amplitude or standard deviation. Once the effects of environmental changes are compensated, the remaining changes in the measured response are a direct result of changes in the structural state. Normalization is closely related to calibration. It is defined as the transformation of sensor output to a nominal value based on a known input and specified environmental conditions. Calibration does not, however, account for any effects the environment may have on a structure’s dynamics.

b) Feature Extraction: It is the process of computing metrics from sensor signals that have the potential to discriminate among the structural states to be identified. Desirable features are ones that are responsive to the structural damage states, yet insensitive to other factors. In many situations, the features are generated from simulation analysis. c)

Dimensionality Reduction: To simplify the problem, only a selected number of features are used. The feature selection is the process of finding a subset of the original features. It can be categorized into filter methods and wrapper


11

methods. In a filter method, relevant features are determined solely based on attributes computed from the data. But, for the wrapper approaches, it is determined based on how well a subset of features performs when used in a classifier. Another method of dimensional reduction is to project the original feature space into a lower dimensional space. Principal Component Analysis (PCA) is commonly used for that. It is the optimum projection in the sense of capturing the maximum data variance for any specified number of basis vectors. The Collaborative Damage Event Detection (CBED) method is quite useful in SHM [22]. In this method, each note continuously collects measured responses and checks the existence of any abrupt changes, which is assumed to be the indication of damage. In the case of abrupt responses, it communicates with neighbors to confirm the existence of damage. Usually damage on structures is accumulated slowly and doesn’t incur abrupt changes. It is also difficult to distinguish the change caused by the real damage from the one caused by environmental conditions (noise or input change) and is prone to giving false positive alarms.

4 Conclusions This paper has reviewed some technical literature on the subject of sensors and technologies used for monitoring the health of structures. The subject is gaining importance in recent times and new sensors cum technologies are reported continuously. It is a very demanding and complex area and a lot of issues are to be considered for a perfect solution. The fabricated sensor systems should be able to inspect or measure without doing any harm or damage of the system. They should also be robust to poor signal-to-noise ratio compared to the level of damage they are trying to detect in the critical infrastructure. Finally, they need to be highly reliable and operate without input for long periods of time, potentially over years. It is expected that the interest in this field is ensured by the constant supply of emerging modalities, techniques and engineering solutions, as well as an increasing need from aging structures, many of the basic concepts and strategies have already matured and now offer opportunities to build upon.

References [1] Choi, H., Choi, S., Cha, H.: Structural Health Monitoring System Based on Strain Gauge Enabled Wireless Sensor Node. In: Proceedings of the Fifth International Conference on Netwroked Sensing Systems, Kanazawa, Japan, June 17-19, pp. 211–214 (2008) [2] Maenaka, K.: MEMS Inertial Sensors and Their Applications. In: Proceedings of the Fifth International Conference on Netwroked Sensing Systems, Kanazawa, Japan, June 17-19, pp. 71–73 (2008)

12


[3] Baptista, F.G., Filho, J.V.: ‘A New Impedance Measurement System for PZT-Based Structural Health Monitoring. IEEE Transactions on Instrumentation and Measurements 58(10), 3602–3608 (2009) [4] Foote, P.D., Read, I.: Optical Sensors for Aerospace Structural Monitoring. IEE Colloquium on Optical Techniques for Structural Monitoring, 2/1–2/6 (April 28, 1995) [5] Kim, D.H., Feng, M.Q.: Real-Time Structural Health Monitoring Using a Novel Fiber-Optic Accelerometer System. IEEE Sensors Journal 7(4), 536–543 (2007) [6] Michie, W.C., Thursby, G., Walsh, D., Culshaw, B., Konstantaki, M.: Distributed Sensing of Physical and Chemical Parameters for Structural Monitoring. IEE Colloquium on Optical Techniques for Structural Monitoring , 9/1–9/6 (April 28, 1995) [7] Bo-lin, S., Bi-feng, S., Fei, C.: New Sensors Technologies in Aircraft Structural Health Monitoring. In: Proceedings of the 2008 International Conference on Condition Monitoring and Diagnosis, Beijing, China, April 21-24, pp. 1–4 (2008) [8] Niu, J., Deng, Z., Zhou, F., Cao, Z., Liu, Z., Zhu, F.: A Structural Health Monitoring System Using Wireless Sensor Network”. In: Proceedings of the 5th International Conference on Wireless Communication, Networking and Mobile Computing, Beijing, China, September 24-26, pp. 1–4 (2009) [9] Anastasi, G., Lo Re, G., Ortolani, M.: WSNs for Structural Health Monitoring of Historical Buildings. In: Proceedings of the 2nd Human Systems Interactions 2009, Catania, Italy, May 21-23, pp. 574–579 (2009) [10] Hedley, M., Hoschke, N., Johnson, M., Lewis, C., Murdoch, A., Price, D., Prokopenko, M., Scott, A., Wang, P., Farmer, A.: Sensor Network for Structural Health Monitoring. In: Proceedings of the 2004 Intelligent Sensors, Sensor Networks and Information Processing Conference, Melbourne, Australia, December 14-17, pp. 361–366 (2004) [11] Wang, D.H., Liao, W.H.: Instrumentation of a Wireless Transmission System for Health Monitoring of Large Infrastructures. In: Proceedings of the 2001 IEEE Instrumentation and Measurements Technology Conference 2001, Budapest, Hungary, May 21-23, pp. 634–639 (2001) [12] Sazonov, E., Li, H., Curry, D., Pillay, P.: Self-Powered Sensors for Monitoring of Highway Bridges. IEEE Sensors Journal 9(11), 1422–1429 (2009) [13] Whelan, M.J., Gangone, M.V., Janoyan, K.D.: Highway Bridge Assessment Using an Adaptive Real-Time Wireless Sensor Network. IEEE Sensors Journal 9(11), 1405– 1413 (2009) [14] Merlino, P., Abramo, A.: An Integrated Sensing/Communication Architecture for Structural Health Monitoring. IEEE Sensors Journal 9(11), 1397–1404 (2009) [15] Chin, J.C., Rautenberg, J.F., Ma, C.Y.T., Pujol, S., Yau, D.K.Y.: An Experimental Low-Cost, Low-Data-Rate Rapid Structural Assessment Network. IEEE Sensors Journal 9(11), 1361–1369 (2009) [16] da Costa Antunes, P.F., et al.: Optical Fiber Accelerometer System for Structural Dynamic Monitoring. IEEE Sensors Journal 9(11), 1347–1354 (2009) [17] Kerrouche, A., Boyle, W.J.O., Sun, T., Grattan, K.T.V., Schmidt, J.W., Taljsten, B.: Strain Measurement Using Embedded Fiber Bragg Grating Sensors Inside an Anchored Carbon Fiber Polymer Reinforcement Prestressing Rod for Structural Monitoring. IEEE Sensors Journal 9(11), 1456–1461 (2009)


13

[18] Grattan, S.K.T., Taylor, S.E., Sun, T., Basheer, P.A.M., Grattan, K.T.V.: In Situ Cross-Calibration of In-Fiber Bragg Grating and Electrical Resistance Strain Gauges for Structural Monitoring Using an Extensometer. IEEE Sensors Journal 9(11), 1355–1360 (2009) [19] Zhang, Y., Tsai, J., Li, G.P., Feng, M.Q., Bachman, M.: MOEMS Acoustic Sensors for Structural Health Monitoring. In: Proceedings of IEEE Sensors Conference, Waikaloa, Hawaii, November 1-4, pp. 2311–2315 (2010) [20] Dondi, D., Di Pompeo, A., Tenti, C., Simunić Rosing, T.: Shimmer: a Wireless Harvesting Embedded System for Active Ultrasonic Structural Health Monitoring. In: Proceedings of IEEE Sensors Conference, Waikaloa, Hawaii, November 1-4, pp. 2325–2328 (2010) [21] Matsushiba, Y., Nishi, H.: Structural Health Monitoring System Considering Faulttolerance. In: Proceedings of the IEEE International Conference on Industrial Informatics (INDIN 2008), Daejeon, Korea, July 13-16, pp. 330–335 (2008) [22] Matsushiba, Y., Nishi, H.: Effective Data Sharing System for Fault Tolerant Structural Health Monitoring System. In: Proceedings of the 11th IEEE International Workshop on Advanced Motion Control, Nagaoka, Japan, March 21-24, pp. 76–81 (2010) [23] Liu, X., Cao, J., Xu, Y., Wu, H., Liu, Y.: A Multi-Scale Strategy in Wireless Sensor Networks for Structural Health Monitoring. In: Proceedings of the 5th International Conference on Intelligent Sensors, Sensor Networks and Information Processing, Melbourne, Australia, December 7-10, pp. 361–366 (2009) [24] DeSimio, M.P., Olson, S.E., Montes de Oca, J.A., Derriso, M.M.: Signal Processing Issues Related to Structural Health Monitoring. In: Proceedings of the 41st Asilomar Conference on Signals, Systems and Computers, Pacific Grove, California, USA, November 4-7, pp. 843–847 (2007) [25] Dezfouli, S., Zabihollah, A.: Structural Health Monitoring of Buried Pipelines under Static Dislocation and Vibration. In: 2010 IEEE/ASME International Conference on Mechatronics and Embedded Systems and Applications, July 15-17. IEEE Computer Society Press, Los Alamitos (2010), doi:10.1109/MESA.2010.5552052. [26] Shrive, N.G.: Intelligent Structural Health Monitoring: a Civil Engineering Perspective. In: Proceedings of the 2005 IEEE International Conference on Systems, Man and Cybernetics, October 10-12, pp. 1973–1977. IEEE Computer Society Press, Los Alamitos (2005), IEEE DOI: 10.1109/ICSNC.2005.1571436 [27] Musiani, D., Lin, K., Simunic Rosing, T.: Active Sensing Platform for Wireless Structural Health Monitoring. In: Proceedings of the 2007 International Conference on Information Processing in Sensors Network, ISNP 2007, Cambridge, Massachusetts, USA, April 25-27, pp. 390–399 (2007) [28] Arms, S.W., Galbreath, J.H., Townsend, C.P., Churchill, D.L., Corneau, B., Ketchan, R.P., Pham, N.: Energy Harvesting Wireless Sensors and Networked Timing Synchronization for Aircraft Structural Health Monitoring. In: Proceedings of the 1st International Conference on Wireless Communication, Vehicular Technology, Information Theory and Aerospace & Electronic Systems Technology (Wireless VITAE), Aalborg, Denmark, May 17-20, pp. 16–20 (2009) [29] Kim, J., Grisso, B.L., Ha, D.S., Inman, D.J.: An All-Digital Low-Power Structural Health Monitoring System. In: Proceedings of the 2007 IEEE Conference on Technologies for Homeland Security, Woburn, MA, USA, May 16-17, pp. 123–128 (2007)

14


[30] Ravinagarajan, A., Dondi, D., Simunic Rosing, T.: DVFS Based Task Scheduling in a Harvesting WSN for Structural Health Monitoring. In: Proceedings of the 2010 Conference on Design, Automation and Test in Europe, Dresden, Germany, March 8-12, pp. 1518–1523 (2010) [31] Furse, C., Smith, P., Diamond, M.: Feasibility of Reflectometry for Nondestructive Evaluation of Prestressed Concrete Anchors. IEEE Sensors Journal 9(11), 1322–1329 (2009) [32] Wijetunge, S., Gunawardana, U., Liyanapathirana, R.: Wireless Sensors networks for Structural Health Monitoring: Considerations for Communication Protocol Design. In: Proceedings of the 17th International Conference on telecommunication, Doga, Qatar, April 4-7, pp. 694–699 (2010) [33] Harms, T., Sedigh, S., Bastianini, F.: Structural Health Monitoring of Bridges Using Wireless Sensor Networks. IEEE Instrumentation and Measurement Magazine, 14–18 (December 2010) [34] Yu, L., Zhu, J.H., Chen, L.J.: Parametric Study on PCA-based Algorithm for Structural Health Monitoring. In: Proceedings of the 2010 Prognostic & System Health Management Conference (PHM 2010), Macau, January 12-14 (2010) paper: MU3072 [35] Yu, L., Xu, P.: An ACO-based Algorithm for Structural Health Monitoring. In: Proceedings of the 2010 Prognostic & System Health Management Conference (PHM 2010), Macau, January 12-14 (2010) paper: MU3053 [36] Liehr, S., Lenke, P., Wendt, M., Krebber, K., Seeger, M., Thiele, E., Metschies, H., Gebreselassie, B., Munich, J.C.: Polymer Optical Fiber Sensors for Distributed Strain Measurement and Application in Structural Health Monitoring. IEEE Sensors Journal 9(11), 1330–1338 (2009) [37] Hann, C.E., Singh-Levett, I., Deam, B.L., Mander, J.B., Chase, J.G.: Real-Time System Identification of a Nonleaner Four-Story Steel Frame Structure – Application to Structural Health Monitoring. IEEE Sensors Journal 9(11), 1339–1346 (2009) [38] Ledeczi, A., Hay, T., Volgyesi, P., Hay, D.R., Nadas, A., Jayaraman, S.: Wireless Acoustic Emission Sensor Network for Structural Health Monitoring. IEEE Sensors Journal 9(11), 1370–1377 (2009) [39] Thomson, D.J., Card, D., Bridges, G.E.: RF Cavity Passive Wireless Sensors with Time-Domain Gating-Based Interrogation for SHM of Civil Structures. IEEE Sensors Journal 9(11), 1430–1438 (2009) [40] Friedrich, M., Dobie, G., Chan, C.C., Pierce, S.G., Galbraith, W., Marshall, S., Hayword, G.: Miniature Mobile Sensor Platforms for Condition Monitoring of Structures. IEEE Sensors Journal 9(11), 1439–1448 (2009)

Self-sustaining Wireless Acoustic Emission Sensor System for Bridge Monitoring Ákos Lédeczi1, Péter Völgyesi1, Eric Barth1, András Nádas1, Alexander Pedchenko1, Thomas Hay2, and Subash Jayaraman2 1

Vanderbilt University, Nashville, TN, USA [email protected] 2 Waves in Solids LLC State College, PA, USA [email protected]

Abstract. A novel approach to structural monitoring of bridges is presented. Acoustic emission sensing has been constrained to hardwired systems up till now because the processing of high bandwidth sensor data on multiple channels requires a lot of energy. The presented prototype wireless system for the real-time detection of active fatigue cracks in bridges overcomes this problem by utilizing a low-power Flash FPGA for signal processing, a novel vibration energy harvester and a sophisticated sleep scheduler.

1 Introduction There are close to 33,000 steel railroad bridges and 600,000 highway bridges in the United States, over 30% of which are structurally deficient or functionally obsolete [1]. Currently, highway bridges in the U.S. are mostly inspected visually [2]. Such inspection only detects an estimated 3.9 percent of existing fatigue cracks [3] Most bridges that people utilize every day were built at least 50 years ago, and were not designed to withstand today’s demanding traffic loads. One of the most comprehensive bridge testing methods is based on acoustic emissions (AE). AE are the stress waves produced by the sudden internal stress redistribution of the materials caused by the changes in the internal structure. Possible causes include crack initiation and growth, crack opening and closure, and dislocation movement among others. Most of the sources of AE are damagerelated; thus, the detection and monitoring of these emissions are commonly used to predict material failure [13]. When load is applied to a fracture-critical member with an existing fatigue crack, stress concentration around the crack tip can eventually cause brittle fracture. Acoustic emission is generated from growing fatigue cracks when a fracture-critical member is stressed. The growing fatigue crack generates a stress wave that travels through the member and can be detected by an AE sensor as well as accurately localized by multiple such sensors using a Time Difference of Arrival (TDOA) method. S.C. Mukhopadhyay (Ed.): New Developments in Sensing Technology for SHM, LNEE 96, pp. 15–39. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

16

Á. Lédeczi et al.

Fig. 1 Wired Bridge Testing

Current AE measurement systems based on hardwired centralized data collection are expensive, power hungry and heavy. This is because AE are high frequency signals and the speed of sound in steel is over 3000 m/s. Hence, detection, feature extraction and accurate source localization requires over 1 MHz sampling rates on multiple channels. Figure 1 shows a typical data collection unit during an inspection. The system is powered by a 1000 Watt power generator and requires long cables (30-200ft) to reach the fracture-critical bridge components where the AE sensors are installed. Just setting up the system takes several hours. Moreover, the expensive equipment needs to be protected against theft and vandalism. Today these systems are used for the occasional inspection of a few selected bridges. However, the importance of real time detection of damage in a critical structural component at an early stage was recently evident in the case of the Oakland Bay Bridge. During retrofitting of the bridge for improved seismic performance, an unexpected crack of dangerous proportions was accidentally discovered on September 7, 2009, prompting complete closure of the bridge due to safety reasons. The bridge could be opened to traffic only after the unscheduled expensive replacement of the affected component was completed. Timely detection of the damage before assuming dangerous proportions would have allowed convenient scheduling of strengthening or replacement of the affected component at significantly lesser cost. This event underscores the obvious need for a cost-effective system that is able to continuously and autonomously monitor the structural health of bridges. The current state-of-the-art is illustrated by the replacement of the collapsed I35W Mississippi River Bridge in Minneapolis, Minnesota. The new bridge has an integrated monitoring system that is completely wired both for power and communication [4]. The main reason a current system must be wired is the lack of

Self-sustaining Wireless Acoustic Emission Sensor System for Bridge Monitoring

17

self sustainability. Current energy harvesting and power management techniques are insufficient to sustain prolonged sensing and monitoring especially if wireless communication is utilized which is notoriously power-hungry. While the cost of a wired system is negligible compared to that of a new bridge, retrofitting the thousands of existing bridges with miles of cables each is cost-prohibitive. The main requirements for a bridge monitoring system then are as follows. A single structure needs to be instrumented with a high number of sensors to cover all critical components. All system components need to rely on energy harvesting since the availability of power near the sensors cannot be assumed and long-term unsupervised operation is desired. Sensor nodes distributed across the bridge need to communicate wirelessly since wiring long bridges is not economical. The system needs to be able to communicate with a central monitoring center to report on the status of the bridge. Finally, the system needs to be low cost to enable instrumentation of a large number of bridges. The remainder of this paper presents the prototype system that meets these requirements enabling the continuous monitoring of railways bridges. We also present out ongoing work toward extending the technology to highway bridges. The next sections will provide a detailed description of the sensor node requirements and design. This is followed by a short description of the signal processing algorithms. The main enabler of the extended operation of the system is a novel vibration energy harvesting technique that is introduced in the subsequent section. Then an overview of the overall system architecture and operation is presented, followed by a summary of the wireless network protocols and offline data evaluation methodology. The paper then concludes with initial test results and our planned future work.

2 Sensor Node A fatigue crack generates an acoustic emission event through the rapid release of elastic energy with each step in the crack growth process. To measure such a phenomenon, an array of three or more acoustic emission sensors are placed on the fracture critical bridge member. Measuring the time of arrival of AE events then enables the localization of the active flaw within the array using standard Time Difference of Arrival (TDoA) techniques. Acoustic emission signal features may be used to estimate fatigue crack growth rates. Useful AE signal features include amplitude, rise-time, counts, signal duration, and energy. Since an AE event is very short (< 1 msec), it contains a substantial amount of frequency information to assist in signal interpretation, and the speed of sound in steel is can range from 2500 to 5850 m/sec, the AE signal needs to be sampled at a high rate, typically at 2-3 MHz to enable feature extraction and source localization. Sampling and especially signal processing at this rate on multiple channels is the most significant design driver of a battery-powered device. Low-power DSP chips simply do not have the horse power needed in this application. In fact, the only choice that can meet the requirements is an FPGAbased solution. However, FPGAs are not low-power thus mandating some kind of power management. Traditional FPGAs do not support sleep modes; if you power

18

Á. Lédeczi et al.

them down, you lose all configuration information. Also, initializing an FPGA incurs a current spike. However, the new flash-based FPGAs overcome these limitations. In sleep mode, their power consumption is practically zero, yet they can wake up in 1 microsecond while conserving their configuration and state. Furthermore, the power consumption in active mode is also reduced significantly. On the down side, flash-based FPGAs are not as powerful as state of the art SRAM-based FPGAs To support the continuous operation of the node for months as opposed to hours on a single charge, the device needs to sleep most of the time. The question is then how and when to wake it up? Using one of the AE channels as a “sentry” is suboptimal since it is power hungry. This could be improved by having one such channel per system as opposed to per node, and alternate the sentry nodes, but then a relatively complex scheme needs to be implemented that wakes up the rest of the nodes using the radio. Again, the radios cannot be listening continuously either, since that is also notoriously power hungry even with low-power radios. Furthermore, having an AE event on one channel (of one node) may mean that we have already missed the same event on the other channels of the same node (and possibly all other nodes). Instead, we employ an application-specific method that fits railway bridge monitoring especially well. AE events are only expected when a train passes over the bridge. If we can detect an oncoming train in time, we can wake up the network. As the train enters the bridge, stresses shall be applied to the bridge’s fracture critical members. A strain sensor, therefore, will detect a train approaching the sensor network provided the strain is of sufficient magnitude. In our prototype, a strain gage channel is employed that is slowly sampled (typically at 1 Hz). The strain data are checked by a low-power duty-cycled microcontroller and if elevated values are observed, it wakes up the rest of the board including the AE channels and the FPGA. While this technique works well for railway bridges, it is not applicable to highway bridges with almost continuous traffic. We are working on an adaptive algorithm that will decide when to turn on the AE channels based on a correlation of past AE events with other sensor modalities. It will try to listen to the AE channels as much as possible given the current battery charge level and the recharge rate from the energy harvester. When significant AE events are observed, it will store other sensor values that preceded these events. The algorithm will be then able to recognize common patterns and continuously adapt its triggering mechanism for the AE channels. AE sensors come in two flavors: with or without pre-amplification. Preamplification is very useful since the typical AE signals are just a few microvolts. However, the preamps need excitation voltage 5V or higher. To conserve power, we opted to support AE sensors with no pre-amplifiers. To compensate for the missing sensor gain, this mandates programmable gain on the node with up to 100 dB amplification. In addition to AE sensing and processing, the second critical requirement of the sensor node is wireless communication capability. The critical design drivers are again low-power and low cost. Many Wireless Sensor Networks (WSN)


19

applications have similar requirements. A few examples include military applications [6], environmental monitoring [7] or precision agriculture [8]. The most widely adopted hardware platform in this domain is IEEE 802.15.4 compliant radio nodes operating in the 2.4GHz range. Typical products provide a few hundred meter communication range, 50 mW power consumption and 250 kbps raw data rate. Sensor Node Architecture The first generation prototype design is illustrated in Figure 2. The four AE channels can be sampled at up to 3 MHz each. However, the resolution of the selected ADC is 12 bits, not the traditional 16 bits. At this sampling rate, higher resolution would have meant a much more complex ADC chip making the board more complicated and expensive. Data from the strain gage is used to correlate stress on the monitored structure to AE from an active fatigue crack, hence, it is connected to the FPGA. The sampling rate is a fixed 100 Hz and a 16-bit ADC is utilized. The channel has a fixed low-pass filter for anti-aliasing. The channel’s equally important task is to wake up the board as trains approach, so the analog

Fig. 2 AEPod Architecture

20

Á. Lédeczi et al.

signal is also connected to one of the input channels of the microcontroller on the wireless communication module. The excitation voltage level is tunable to decrease the necessary power in sleep mode. The on-board SD card and the provided USB connectivity increase the flexibility of the node tremendously. The SD card can store the measurements for a long deployment; it can be swapped out without having to download potentially large amounts of data. The USB provides fast data downloads and is also used to charge the battery. It is connected through the FPGA and not the microcontroller to support the highest data rates. The on-board SRAM memory can be utilized for short-term waveform storage, for example. Low power and high intensity LEDs are used to display basic status information. A real time clock is included to correlate measurements with train schedule. The battery selected is a 10 Ah lithium ion with a small form factor. A switching regulator with high efficiency is utilized. Remaining battery capacity is monitored with a coulomb counter. Figure 3 shows the prototype board. Table 1 summarizes the current draw of the AEPod at 3 V in active mode. Considering the utilized 10 Ah battery, the board supports approximately 80 hours of active mode. Note that these numbers do not include any wireless data download. The data for sleep mode (with duty cycling the microcontroller to monitor the strain channel) is summarized in Table 2. The 5mA draw means that the board can sustain 2000 hours of sleep mode. The actual lifetime will depend on the ratio of active to passive mode, i.e. the frequency of trains passing through the bridge. For example, if the node is active 1 hour per day, the expected lifetime of the node on a single charge is about 6 weeks.

Fig. 3 First Generation Prototype


21

Table 1 Current draw in active mode

Component 4x AE channel FPGA Communication module with radio off Communication module with radio on Strain Gage channel Other components/loss Power conversion efficiency

mA 48 35 6 25 3 3 80%

Total (no radio): Total (with radio):

~125 ~150

Table 2 Current draw in sleep mode

Component 4x AE channel FPGA Comm module (with duty cycling) Strain Gage channel (lower voltage) Other components/loss Power conversion efficiency Total:

mA 0 0 1 1 2 80% 5

Sensor Node Revision The prototype system proved that the flash-based FPGA technology enables high sampling rate with multiple signal streams and powerful node level data processing without significant impact on the power budget. It also demonstrated that duty cycling with smart wake-up triggers are essential for taking advantage of the benefits of the new platform (extremely fast wake-up and low static power consumption). Based on our experience with the prototype hardware platform and taking into account additional requirements and opportunities, we have redesigned the platform. The changes add new sensor modalities to enable this platform to be used in a much wider range of applications and to provide the necessary foundation for advanced duty cycling and triggering mechanisms to be developed. Figure 4 shows the architecture of the revised sensor node. We opted for a slightly different integrated module (Meshnetics ZigBit Amp [9]), which is code compatible with the originally used Crosbow IRIS module,

22

Á. Lédeczi et al.

but has an integrated 20dB RF amplifier providing significantly higher communication ranges eliminating many deployment constraints. The prototype system was designed to be deployed and operational for several weeks autonomously. Now we are aiming at much longer periods between service events or any human interactions with the nodes. For this, we extended the power management subsystem with energy harvesting options. The sensor platform will support renewable energy sources with special emphasis on vibration-based energy harvesting. Other related projects [10][11] were built around MEMS accelerometers as the single sensing source. These sensors are widely available, inexpensive and the power requirements of the sensor itself and that of the signal processing tasks are low, making this approach feasible for wireless sensor networks. We incorporated 3D MEMS accelerometers in the revised sensor platform. We also extended the prototype with additional low-frequency high-precision strain gauge channels, since these are excellent candidates for application specific wake-up triggers. Based on these sensor modalities, we will be able to use the same sensor platform in a wide range of applications where vibration, stress, ultrasonic elastic waves are good indicators of the health of the physical structure.

Fig. 4 Revised Sensor Node Design


23

Signal Processing The objective of interpretation in acoustic emission bridge inspection is to assess the significance of sources of emission. The general approach begins with filtering out non-relevant emissions. A variety of acoustic signals occurs during the monitoring process, such as noise from loose fasteners and working members, mechanical noise transferred from the track and random electrical signals. These may mask the acoustic signals from a growing crack. Since an AE event from a growing crack has a typical, characteristic waveform, signal characteristics (features) are used to identify and filter crack-related acoustic events from noise. AE data are evaluated in terms of activity and intensity. Activity is defined in terms of acoustic events that are detected inside the sensor array by all four sensors. Intensity is defined as the average signal strength of the acoustic events in dB: • Activity: The number of events that occur within the sensing array. For an acoustic source to be classified as an event, it must be picked up by all four sensors and originate from inside the array. Activity may be classified as Critically Active if events are observed consistently at peak load, Active if randomly observed over the load spectrum, and Inactive. • Intensity: The average amplitude, in dB, of the events. Acoustic emission may be classified as Low Intensity (< 50 dB), Intense (50 - 75 dB), and Critically Intense (> 75 dB). Based on AE activity and intensity, the AE source index is developed as shown in Figure 5. Activity and intensity metrics are derived from the features of the AE signals and their dependence on the load used to stimulate AE. Feature extraction and source localization are carried out in the FPGA. The frequency response of the AE sensors themselves provides sufficient filtering [17] to enable a relatively simple time-domain based signal processing approach. This fits the problem well since the relevant AE signal features are all time-domain parameters. The configurable hardware implementation enables signal processing in a streaming manner at the sampling frequency; each sample is fully processed before the next one arrives. That is, there is no need to buffer the signals.

Fig. 5 AE Classification

24

Á. Lédeczi et al.

The most important task of the processing core is to detect and measure the time of arrival of the acoustic events and to extract its characteristic features. Last but not least, the detection logic should work reliably with a few parameters to minimize the calibration requirements of the system. Figure 6 shows the time series of a typical acoustic event along with the current operational parameters (letters) and the extracted features (numbers). The detection is triggered by a simple threshold (A) crossing condition. The start of the event (1) is the time instant of this threshold crossing. The end of the acoustic emission is identified if the signal does not cross the threshold level for longer than a timeout parameter (B). Between these two events the signal is monitored and the following features are calculated: the number of threshold crossings (using hysteresis to eliminate the effects of noise or riding waves) (5) – this is a good estimate of the fundamental frequency of the signal, maximum amplitude (4), rise time (3) – the elapsed time between the initial threshold crossing and the time when the signal reaches the maximum amplitude, the length of the event (2) – the time span between the initial and last threshold crossings and the energy of the signal (6) – sum of the squared sample values. Events on multiple channels are then evaluated. If at least three channels detect AE, then the TDoA values are checked for consistency. If the difference between any two time stamps is larger than it takes for the sound to travel the distance between the corresponding sensors, the event could not have come from the same source. Such inconsistent observations are discarded. If the timestamps are consistent, then the source is localized using the standard TDoA equations. If the AE did not originate from within the sensor array, it is also discarded.

Fig. 6 AE Event Features

3 Energy Harvesting The goal is to make the system self-sufficient from a power standpoint by harvesting kinetic energy from the structural vibrations of the bridge at, or very


25

near, the location of each acoustic emission sensor node to provide power for its on-board sensing, processing and wireless communication functions. Given that the power harvesting for each node will be done locally and not in a centralized location, it is required to design an energy harvesting module that can accept and adapt to a broad range of excitation frequencies. The excitation frequency will vary by bridge, location of the sensor on the bridge, and traffic load on the bridge. The goal is to therefore design a modular energy harvester that will extract vibrational kinetic energy for a changing primary vibration mode by optimally changing its natural frequency. Energy harvesters have been a growing topic of research interest and show their greatest potential for wireless sensor node applications. In particular, a survey of recent work on energy harvesters shows the greatest effectiveness for devices in the frequency range of 10-100 Hz [18]. Of the types that harvest energy from motion, three transduction methods are generally available: 1) electromagnetic, 2) electrostatic, and 3) piezoelectric. Electromagnetic transduction, in either a linear device or a rotary device, is generally not well suited for MEMS scale devices given the challenge of manufacturing coil windings and the integration of permanent magnets. Electromagnetic transduction is however practical at more macroscopic scales such as that under consideration here. Electrostatic transduction is impractical and inefficient at macroscopic scales, suitable only for the microscale [18]. Piezoelectric materials such as PZT (lead zirconate titanate) require a mechanical transmission system for scales larger than MEMS, require high voltage power electronics and are typically not well impedance matched for mechanical vibrations characterized by the amplitudes and frequencies found in a bridge structure. Of the electromagnetic transducer variety of energy harvester, most utilize a proof mass suspended by a mechanical spring such that the mass-spring system’s resonant frequency (damped natural frequency) is tuned to the largest frequency component of the excitation [19]. Commercial incarnations of this concept appear in such applications as self-winding watches and self-powered flashlights. One such shake-powered flashlight contains a 150 gram energy harvester and is able to produce 200 mW when excited at its resonant frequency of 3.3 Hz [19]. The down side of such devices is that they must be excited near their resonant frequency to generate an appreciable amount of power. A recent effort to design an energy harvester for a bridge monitoring system resulted in a device with a resonant frequency of 3.12 Hz and an energy extraction mechanism that behaved as a damper [20]. Such a system, where only the damping behavior is controlled, has relatively little ability to tune the resonant peak. Unfortunately this is typical of the research to date for such applications. For bridge monitoring, the ability to alter the resonant frequency of the harvester to match that of the excitation frequency is imperative. Modern highway bridges posses a fundamental frequency typically in the range of 2-5 Hz, and very stiff bridges have fundamental frequencies in the range 10-15 Hz. [21]. This range is dependent on the design of the bridge, but it is also dependent on the traffic load on the bridge. The excitation frequency imparted to the harvester can therefore vary for a given bridge and even for a given time of day. For this primary reason,

26

Á. Lédeczi et al.

an energy harvester needs to be developed that has the novel ability of altering its damped natural frequency. A secondary reason to design for such an ability is to promote modularity in sensor node systems for bridges. In this way, one common sensor node system can be utilized for a wide variety of bridges. The energy harvester concept presented here offers a novel method for altering the resonant frequency by one or two orders of magnitude in real-time. To the authors’ knowledge, such an approach does not currently exist. Resonant Peak Shifting via Control The novel harvester contains an inertial element vertically suspended in parallel by a spring and a linear generator. Alternate designs with a cantilevered mass and typical motor/generators are also possible. The extraction of vibrational kinetic energy can be accomplished through a combined design of mechanical and control elements within the harvester. The mechanical mass-spring portion of the harvester offers a device with a natural frequency designed to be in the center of the expected excitation frequency range of all bridges intended for monitoring. The control design and resulting control action results in a dynamic behavior of the linear generator that mimics a damper, a spring, and a mass. The damping behavior will be used for power extraction. The spring and inertial behaviors will require that the linear generator also act as a linear motor, and will require regenerative electronics. This spring behavior will add to, or subtract from, the influence of the mechanical spring to allow a broad range of frequency shifting of the system’s overall damped natural frequency. Likewise, the mimicked inertial behavior can be used to alter the damped natural frequency. In this manner, the device will be able, in real-time, to adjust its resonant response so as to capture more energy than with a fixed resonant frequency harvester. Referring to Figure 7, consider an energy harvester with a proof mass M and a mechanical spring with stiffness k. Declaring y = 0 at the static hung length of the spring with the proof mass attached under the influence of gravity, the following equation relates the absolute motion y of the proof mass to the excitation motion x as influenced by the spring force and the controllable force imparted by the voicecoil (linear generator/motor) u: M y = u − k ( y − x) . (1) Conventional approaches essentially dictate the voicecoil to behave as a damper acting on the relative motion of the base and the proof mass by utilizing a control law such as, u = −bc ( y − x ) .

(2)

Consider instead a control law that requires the voicecoil to behave as a damper, a spring and an inertia through the measurement and feedback of acceleration, relative velocity and relative position, u = −M c y − bc ( y − x ) − k c ( y − x) .

(3)

The transfer function of the closed-loop system relating the proof mass motion to the excitation motion therefore appears as,

Self-sustaining Wireless Acoustic Emission Sensor System for Bridge Monitoring ( kc + k ) bc 2ξω n s + ω n2 Y (s ) (M +M ) s + (Mc + M ) = 2 c bc = 2 . ( kc + k ) X ( s ) s + ( M c + M ) s + ( M c + M ) s + 2ξω n s + ω n2

27

(4)

From Equation (4) it can be seen that the natural frequency of the harvester is ω n = (k c + k ) /( M c + M ) . It can also be seen that for the conventional control law of Equation (2) lacking a spring-like or inertia-like behavior, the natural frequency of the harvester cannot be changed. The resulting damped natural frequency of the system can be changed only slightly. In contrast, by including a spring-like contribution and/or an inertial-like contribution to the system through the use of a non-zero value for kc and M c in Equation (3), it allows us to influence the natural frequency of the system and the resulting damped natural frequency by potentially orders of magnitude. The fact that both kc and M c can be altered in concert extends the range of adaptability of the natural frequency as limited by sensor noise in position or acceleration alone. The control approach also allows us to arbitrarily influence the amplified proof mass motion so as to magnify it as much as possible exceeding the mechanical limits of the device. Power extraction can then be optimized in real-time by optimizing the average power ∫ bc ( y − x ) 2 dt T

dependent on bc and the height of the appropriately shifted resonant peak as influenced by all three specifiable parameters bc , kc and M c .

Fig. 7 Schematic of a vibrational energy harvester

Comparison of Conventional and the New Approach – a Case Study Presented below is a case study comparing the conventional energy extractor control approach to the adaptive resonant peak shifting control approach. For this study, we will restrict our attention to varying the stiffness of the system. Similar conclusions hold for varying the inertial properties of the system. For the case

28

Á. Lédeczi et al.

presented, the proof mass M = 0.227 kg (0.5 lbs), and the mechanical spring stiffness k = 224 N/m (1.28 lbs/inch). Consider the following excitation, x(t ) = A sin(ωt ) (5) where the amplitude is set to unity and the frequency of excitation can take on any value between 0.1 and 20 Hz (corresponding to the expected fundamental frequencies of most bridges under most conditions). For the conventional control approach utilizing the control law of Equation (2), Figure 8a shows a family of proof mass magnitude responses for a range of values for bc . As can be seen, the frequency of the peak is not able to be shifted and results in low vibration amplitude when the excitation departs from the designed mechanical natural frequency of ωn = k / M = 5Hz . 50

40

Magnitude (dB)

30

20

10

0

-10

-20 -2 10

-1

10

0

10 Frequency (Hz)

1

10

2

10

(a) 50

40

Magnitude (dB)

30

20

10

0

-10

-20 -2 10

-1

10

0

10 Frequency (Hz)

1

10

2

10

(b) Fig. 8 Magnitude responses of the (a) proof mass for a range of controllable damping values and (b) proof mass for a range of spring and damping controller parameters


29

For the control approach utilizing the control law of Equation (3), Figure 8b shows a family of proof mass magnitude responses as both kc and bc are allowed to vary. As can be seen, the extra control degree of freedom offered by k c in addition to bc allows two orders of magnitude of resonant peak shifting. This allows a system capable of centering its resonance peak on excitation frequencies from 0.1 Hz to 20 Hz while maintaining a large amplitude of the proof mass for adequate power extraction. Additionally, the control law allows us to arbitrarily set the height of the peak. For the case shown below, a damping ratio of ξ = 0.01 was selected resulting in a magnitude of about 34 dB (an amplification in the proof mass amplitude of 50 times that of the excitation amplitude). The control parameters were set according to the frequency of excitation ω , which can be easily measured by the on-board accelerometer (ADXL330), and the desired damping ratio: kc = Mω 2 − k ,

(6)

bc = 2Mξω .

(7)

For the excitation frequency range of interest that would encompass nearly all bridges and bridge conditions (0.1 Hz to 20 Hz), the resulting range of controlled stiffness is -223.9 N/m to 3360.6 N/m and the resulting controlled damping coefficient range is 0.003 Ns/m to 0.57 ns/m, both well within achievable limits. It should be noted that negative stiffness values that subtract from the mechanical stiffness of the system are possible. It should also be noted that regenerative power electronics are needed to implement the control law; but these are available as offthe-shelf components. For a common bridge frequency of 5 Hz [22], and an amplitude of 1mm (2mm peak-to-peak), the harvester will generate an average of 170 mW at 100% conversion efficiency. At a conservative estimate of 50% harvester conversion efficiency, this results in an average generated power of 85 mW. The maximum power draw from the sensor node is 300 mW. The harvester would therefore enable a 28% duty cycle. On the other hand, an excitation frequency of 10 Hz and a conversion efficiency of 50% result in 670 mW, or enough to power the sensor node all the time. Initial Experimental Results for Resonant Peak Shifting via Control Shown in Figure 9 is the experimental setup used to verify the notion of resonant peak shifting by controlling the overall stiffness of the harvester. The setup consists of two pinned beams. The lower beam represents the bridge motion by using a voice coil to actuate the beam at different frequencies and amplitudes. The upper beam represents the proof mass sprung to the base. The harvester voice coil serves the function of the linear generator/motor and of implementing Equation (3). The pinned beams isolate the motion to one dimension and also allow accurate measurement and scaling of the motions.

30

Á. Lédeczi et al.

Fig. 9 Photograph of the experimental setup

Figure 10 shows amplitude data taken from the experimental setup. Two cases are shown: one where the harvester acts only as an energy extractor (solid line), and a second case where the harvester includes controlled spring behavior. As seen in the plot, the harvester has a natural frequency of about 3.4 Hz. By including spring behavior, the harvester is able to augment the physical spring in the system and subsequently shift the resonant peak to 6.4 Hz. This demonstrates that should the excitation frequency vary, the harvester will be able to alter the resonant frequency in order to achieve a large amplitude motion at the new excitation frequency and consequently generate power. The use of regenerative power electronics will be critical in realizing this concept. 20 no virtual stiffness k c with virtual stiffness k

15

Magnitude Y(s)/X(s) (dB)

10 5 0 -5 -10 -15 -20 -25 -30 -1 10

0

10 Excitation Frequency (Hz)

Fig. 10 Experimental resonant peak shifting

1

10

c


31

4 System Architecture and Operation The architecture of overall system is shown in Figure 11. It consists of a number of sensor nodes and a base station. The sensor nodes have piezoelectric AE sensors (for accurate crack localization in 3D), MEMS accelerometers and strain gauges attached to them. Power to the sensor node is provided by the on-board battery and depending on the location of the sensor node and its environment, by a vibration energy harvesting unit that is connected and mounted at a carefully selected nearby position that offers the most vibration. Upon installation, the sensor nodes form an ad-hoc wireless network, monitor the AE, MEMS accelerometer and strain gauge channels and execute local signal processing and sensor fusion tasks and exchange radio messages with other nodes and the base station. The node level signal processing algorithms detect transient events (e.g.: acoustic emissions) and extract their salient features, such as maximum amplitude, duration, total energy, are extracted and the arrival times on the different channels are recorded. Local data fusion tasks filter out inconsistent data and execute preliminary localization and classification steps based on the extracted features. Insignificant events are discarded. Events and their features can be stored onboard and optionally sent to the base station or neighboring nodes. The decision, whether and if so, when to report events and/or involve other nodes in the data fusion, depends on the extracted features and the energy available to the node. We leave the algorithm that will schedule event reporting and distributed data fusion based on the severity of the event and the energy level of the nodes in the network for future work. Note, however, that unlike in delay-tolerant networks [23] critical events will need to be sent to the base station immediately using guaranteed delivery irrespective of the energy levels. These are events that may demand urgent action, such as potentially evacuating the immediate area.

Fig. 11 System Architecture

32

Á. Lédeczi et al.

The base station, an embedded PC class device, has a dual purpose in the system. First, it runs high-level structural health assessment algorithms based on the model of the given structure. Note that this aspect of the system is beyond the scope of this paper. Second, it is equipped with a GSM modem to report to a central monitoring station. Power management Node-level power management is a key component of the system. The sensor nodes have relatively high power consumption due to multichannel signal processing, the required high sampling rates on some of these channels (AE transducers) and node level data fusion and (de)compression. The worst case power consumption of the revised board is estimated at 300mW with every component turned on. Given a 10 Ah battery (3.6V Li-Ion cell), that comes to 120 hours of operation at full power. Even with energy harvesting, we do not expect that the node will be able to sustain continuous operation indefinitely. Although, low-power FPGAs – due to their several clock networks, clock scaling options, deep sleep modes and fast wakeup times – provide an excellent research platform for experimenting with various power-saving schemes, the triggering mechanism driving these modes is a cornerstone issue in any application. Our first generation system solved this problem specifically for railway bridges. There, AE events are only expected when a train passes over the bridge. As the train enters the bridge, stresses are applied to the structure. The strain gage channel is slowly sampled by the duty-cycled microcontroller and if elevated values are observed, it wakes up the rest of the board. The prototype has an estimated lifetime of 6 weeks on a single charge on a typical railway bridge. Clearly, the same strategy is not applicable to highway bridges which have more or less continuous traffic. The revised platform provides a more general framework for various triggering schemes. These triggering mechanisms need to minimize the chance of missing significant sensor events at the lowest possible consumed energy. The power management framework will support low power sampling with periodic or stochastic duty cycling and clock scaling using several clock domains (a clear advantage of using the FPGA platform) and will provide accurate estimates on the current energy budget (charge level, charge rate). The power management framework will also be notified by the sensor monitoring and signal processing layers when significant events are detected, thus it will be able to learn correlations between channels and sensor modalities. This runtime learning capability will augment the static ruleset created by the application developers and domain experts. We are working on a simple adaptive algorithm whereas the node will dynamically adjust its behavior by observing the frequency and severity of sensor events (i.e. cracks) and the correlation among all sensor channels. Consequently, the threshold values in the table will be continuously adjusted based on the behavior of the structure and energy harvesting opportunities. This is well suited for the envisioned very long term deployments, since the behavior of the components is expected to slowly change as the structure ages. The algorithm will


33

follow this trend. Note, however, that the algorithm needs to remain relatively simple, as it will run on a low-power, resource constrained microcontroller. Low power listening When the radio transceiver is turned on, it draws approximately 15mA operating at 3V, which can deplete the batteries quickly. Therefore, in order to extend battery life, radios must duty-cycle, i.e. wake up for short periods while staying turned off the majority of time. The low-power radio stack we implemented for the transceiver works as follows. In receive mode, the transceiver is configured to duty-cycle, waking up periodically to check if a transmission is on the way. If no signal is detected, the transceiver goes back to sleep mode. If there is a transmission in progress, the transceiver stays awake to receive (and optionally acknowledge) the packet. Since packets are often received in bursts, the receiver stays on for a short period of time after a reception to wait for successive packets. When no more packet transmissions are detected, the receiver goes to sleep mode and continues duty cycling. While the power consumption of the receiver can be drastically decreased this way, the transmitter must ensure that the radio channel is modulated long enough for the receiver to detect an incoming message. Therefore, the transmitter must keep the radio channel modulated longer than the sleep interval of the recipient. The longer the receiver’s sleep interval is, the less power is consumed. However, long sleep interval at the receiver mandates that the sender must modulate the channel long enough for the receiver to detect the transmission – which, results in higher power consumption of the transmitter. Since such tradeoffs are application specific, we implemented the low power radio stack such that the receiver’s sleep intervals, as well as the modulation durations are configurable during run-time. This implementation is suitable for a wide range of applications and allows for reconfiguration to adapt to operating conditions (e.g. available battery power) in a flexible way. Other network services The most important distributed service necessary in the system is multi-hop message routing. We utilize the Directed Flood Routing Framework [15] developed previously for our acoustic countersniper system [6]. For source localization we use the four AE sensors attached to the same AEPod sharing the same clock. In the future, we may want to correlate AE events across sensor nodes requiring precision time synchronization. The Routing Integrated Time Synchronization (RITS) protocol [14], also developed for the countersniper system, can provide 1-2 microsecond accuracy per hop. Reliability The FPGA-based architecture also promotes a more robust approach for implementing deeply embedded but still software driven system. It is a well-known problem that traditional (microcontroller-based) embedded systems

34

Á. Lédeczi et al.

are extremely sensitive to firmware/software errors. While many of these errors are similar to those in normal desktop applications, cyber-physical systems also experience more hidden problems due to their long-term deployment [24]. The typical answer for these problems is the use of watchdogs, grenade timers [25]. These “solutions” have their own negative side effects (application state is completely lost, non-trivial transient management on the cyber-physical boundary during reset). Instead of a purely software driven architecture we are using a layered approach, where lower level sensing, basic signal processing and communication tasks are implemented in the FPGA fabric. These hardware-driven tasks can execute autonomously and need software interaction for configuration and higher level services only. By decoupling the more complex and error prone software layer (running on a CPU softcores) from the physical interfaces, a potential failure and/or reset is less visible to the outside world. As an added benefit of this flexible hardware/software boundary, the standard and rudimentary watchdog approach can be enhanced by additional application specific checks enforced by the deterministic hardware layer (sophisticated memory/IO protection, which typically not present in MCUs, temporal rules, duty cycle enforcement). The development of such smart hardware (IP core)-based supervisors results in a rich set of general schemes potentially applicable in other domains as well. Offline Evaluation It has been shown that the AE event rate per cycle is proportional to the crack propagation rate per cycle [24]. In experiments and from the proportionality between events rate per cycle and crack propagation rate per cycle, this study suggested a relation between the observed event count over any cyclic interval, and the crack area created in this interval. These findings are mostly empirical and are obtained experimentally. They allow the determination of the fatigue life curves based on AE test data. Such curves can be derived for a material or structure and provide an assessment of fatigue damage to material containing a crack. At higher stress intensity levels, the yield of emission per unit of crack extension is higher, a consequence of the larger amounts of stored energy available at higher stress intensities. This provides a basis to link acoustic emission with the fracture mechanism and to establish the relationship between emission and stress intensity factor. A Fatigue Assessment Index (FAI) is then defined based on the AE source activity, intensity and the related fatigue crack. The corresponding recommended actions can be applied to each zone ranging from no action required, through various levels of follow up NDE or analysis, up to taking the structure out of service. These recommendations provide bridge engineers with information to plan, schedule and prioritize maintenance or replacement operations.

5 Initial Field Testing The objective of the field test was to benchmark the wireless system against the wired system. The tested component is shown in Figure 12. The sensors connected


35

to the AEPod are marked by the circles in the figure. The four sensors connected to the traditional wired system are shown inside the wireless sensors.

Fig. 12 Field Experiment

The events recorded by the wireless system are shown in Figure 13. In this figure, five different train passes are observed. The events recorded inside the array are superimposed upon the Train 1, 2-3, and 5 strain curves. Note that there was some overlapping of trains 2 and 3 as they passed over opposite tracks. For this analysis, they were treated as a single train. No events were recorded for train 4 by either the wired or wireless systems.

Fig. 13 Detected AE events (circles) superimposed on the strain signal

36

Á. Lédeczi et al.

The events were further analyzed for intensity levels. The wireless system detected a total of 11 events over 5 train passes at an average intensity of 52 dB. For the same 5 trains, the wired system detected 9 events at an average intensity of 54 dB. For train 1, the wireless and wired systems detected 6 and 5 events from within the array. These numbers are close enough to suggest that the two systems are performing comparably since a missed event may be attributed to sensor location, surface preparation, and sensor coupling to the structure. Similarly, the average intensities are comparable. The 4 dB difference, again, is well within the realm of the considerations cited above. The results from trains 2-3 also indicate that the wireless system is detecting the same events the wired system is detecting. Both systems detected 4 events inside the array at comparable intensities ~ 51 dB (wireless) and 58 dB (wired). Train 4 data results suggests further that the wired and wireless systems are performing comparably since neither array detected any events. Finally, the train 5 result shows that the wireless system detected one event that the wired system did not pick up. The result is not indicative of sensitivity differences between the two instruments. It is again suggestive that sensor location, surface preparation, and acoustic coupling will have minor influences on activity and intensity. A location comparison of the events was also carried out. In the bridge component tested, there is commonly a 1-3” location error due its structure and multiple fasteners in the joint. These features affect line-of-sight between the sensor and the source as well as the sound wave velocity, both of which influence accurate source localization. In this test, all events were clustered in an approximate 3” x 2” area. They were located in the vicinity of the crack tip (leading edge of crack) and are within the margin of location error associated with such bridge components.

6 Related Work Wireless structural monitoring has been an active area of research. Most approaches utilize accelerometers and/or strain gages to analyze the vibrating structure. Lynch and Loh present a comprehensive overview of the state of the art in [26]. Acoustic emission testing is a significantly different problem due to the required sampling rate that is typically two orders of magnitude higher than vibration monitoring. Grosse et al. [27] and independently Yoon et al. [28] created a wireless AE sensor node targeted primarily at concrete structures. Both are single channel sensors with maximum sampling rate in the 100 kHz range. Localization using separate sensor nodes mandates time synchronization across the wireless nodes. The current state of the art in time synchronization accuracy on the kind of wireless technology their and our system utilize is worse than a microsecond [14]. This translates to localization error of over 30 cm in steel. The 4 channels on our board share the same physical clock completely eliminating this source of error. Also, the order of magnitude higher sampling rate of our node makes TDoA measurements and hence, source localization much more accurate.


37

7 Conclusions The paper presented a prototype wireless 4-channel acoustic emission sensor node that supports high sampling rates and hence, enables the detection and accurate localization of fatigue cracks. Having a wireless bridge monitoring system has many advantages. First of all, current wired systems were designed for short-term inspection and are not cost-effective or practical for long-term monitoring. Hence, our system enables a revolutionary paradigm shift in infrastructure maintenance. As such, the change will not happen overnight. We expect that our system will be initially used for inspection before it will be widely adopted for the permanent instrumentation of bridges. Even then it will demonstrate significant savings in time, effort and cost not having to deploy long cables in sometimes hard to access areas. Being a first generation prototype, the current system does have some shortcomings. The lifetime of the nodes on a single charge under real life conditions will be probably be two to four weeks depending on the traffic on the bridge. Wireless data download is also limited primarily due to power constraints also. In our latest generation design, we are addressing these problems and extending the capabilities of the node to be able to support monitoring of highways bridges as well. To this end, we presented a novel energy harvester design that utilizes the vibrations of the bridge itself due to wind and traffic. Laboratory experiments show that the amount of energy that can be gathered with it is an order of magnitude higher than what is provided by current piezoelectric designs. Acknowledgement. This research was supported in part by the National Science Foundation awards CNS 0964592 and CNS 1035627 and the U.S. Transportation Research Board.

References 1. Deficient Bridges by State and Highway System, FHWA, http://www.fhwa.dot.gov/bridge/deficient.htm 2. Reliability of Visual Bridge Inspection, Turner-Fairbank Highway Research Center, Federal Highway Administration (March 2001) 3. TechBrief: Reliability of Visual Inspection for Highway Bridges, Turner-Fairbank Highway Research Center, FHWA (September 2001), http://www.tfhrc.gov/hnr20/nde/01105.pdf 4. Bridge Structural Monitoring System, http://www.dataq.com/applicat/articles/ bridge-structural-monitoring.html 5. Issues and Options in Infrastructure Investment, Congressional Budget Office (May 2008) 6. Ledeczi, A., et al.: Countersniper System for Urban Warfare. ACM Transactions on Sensor Networks 1(2), 153–177 (2005)

38

Á. Lédeczi et al.

7. Selavo, L., et al.: LUSTER: wireless sensor network for environmental research. In: Proc. of the 5th International Conference on Embedded Networked Sensor Systems (SenSys), Sydney, Australia (November 2007) 8. Crossbow Inc. eKo System, http://www.xbow.com/eko/index.aspx 9. Meshnetics Zigbit Amp module (2009), http://www.meshnetics.com/zigbee-modules/amp/ 10. Xu, N., Rangwala, S., Chintalapudi, K., Ganesan, D., Broad, A., Govindan, G., Estrin, D.: A Wireless Sensor Network For Structural Monitoring. In: Proceedings of the 2nd ACM Conference on Embedded Network Sensor Systems (SenSys 2004), Baltimore, MD (November 2004) 11. Kim, S., Pakzad, S., Culler, D., Demmel, J., Fenves, G., Glaser, S., Turon, M.: Health Monitoring of Civil Infrastructures Using Wireless Sensor Networks. In: 6th International Symposium on Information Processing in Sensor Networks, IPSN 2007, April 25-27, pp. 254–263 (2007) 12. Hay, T., Jayaraman, S., Ledeczi, A., Volgyesi, P., Al Sammane, G.: Wireless Technology for Steel Bridges.In: Proceedings of the American Society for Nondestructive Testing Conference on Structural Materials and Technology, Oakland, CA (September 2008) 13. Miller, R.K.: Acoustic Emission Testing. In: Nondestructive Testing Handbook, 3rd edn., vol. 6. American Society for Non-destructive Testing, Columbus (2006) 14. Kusy, B., Dutta, P., Levis, P., Maroti, M., Ledeczi, A., Culler, D.: Elapsed Time on Arrival: A simple, versatile, and scalable primitive for canonical time synchronization services. Int. Journal of Ad Hoc and Ubiquitous Computing (2006) 15. Maroti, M.: The Directed Flood Routing Framework. In: Proc. of Fifth ACM/IFIP/USENIX International Conference on Middleware, Toronto, Canada (October 2004) 16. Sinclair, A.C.E., Connors, D.C., Formby, C.L.: Acoustic Emission Analysis during Fatigue Crack Growth in Steel. Journal of Material Science Engineering 28 (1977) 17. Huang, M., Jiang, L., Liaw, P.K., Brooks, C.R., Seelev, R., Klarstrom, D.L.: Using acoustic emission in fatigue and fracture materials research. JOM 50(11), 1–14 (1998) 18. Mitcheson, P.D., Yeatman, E.M., Rao, G.K., Holmes, A.S., Green, T.C.: Energy Harvesting From Human and Machine Motion for Wireless Electronic Devices. Proceedings of the IEEE 96(9), 1457–1486 (2008) 19. Paradiso, J.A., Starner, T.: Energy scavenging for mobile and wireless electronics. IEEE Pervasive Computing 4(1), 18–27 (2005) 20. Li, H., Pillay, P.: A Methodology to Design Linear Generators for Energy Conversion of Ambient Vibrations. In: IEEE Industry Applications Society Annual Meeting, IAS 2008, pp. 1–8 (October 2008) 21. Paultre, P., Chaallal, O., Proulx, J.: Bridge Dynamics and Dynamic Amplification Factors – A Review of Analytical and Experimental Findings. Canadian Journal of Civil Engineering 19, 260–278 (1992) 22. Jozef, M.: Vibration Measurement of Highway Bridges. In: Proceedings of the 40th International Conference Experimental Stress Analysis, Prague, Czech Republic, pp. 169–174 (2002) 23. Juang, P., Oki, H., Wang, Y., Martonosi, M., Peh, L., Rubenstein, D.: Energy-Efficient Computing for Wildlife Tracking: Design Tradeoffs and Early Experiences with ZebraNet. In: Tenth International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS-X (October 2002)


39

24. Huang, Y., Kintala, C., Kolettis, N., Fulton, N.D.: Software rejuvenation: analysis, module and applications. In: Twenty-Fifth International Symposium on Fault-Tolerant Computing, FTCS-25. Digest of Papers, Pasadena, CA, USA, pp. 381–390 (June 1995) 25. Dutta, P., Hui, J., Jeong, J., Kim, S., Sharp, C., Taneja, J., Tolle, G., Whitehouse, K., Culler, D.: Trio: enabling sustainable and scalable outdoor wireless sensor network deployments. In: Proc. 5th International Conference Information Processing Sensor Networks (IPSN 2006), pp. 407–415 (2006) 26. Lynch, J.P., Loh, K.J.: A Summary Review of Wireless Sensors and Sensor Networks for Structural Health Monitoring. The Shock and Vibration Digest 38(2), 91–128 (2006) 27. Grosse, C., McLaskey, G., Bachmaier, S., Glaser, S.D., Krügera, M.: A hybrid wireless sensor network for acoustic emission testing in SHM. In: Proc. of SPIE Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2008, vol. 6932 (2008) 28. Yoon, D., Lee, S., Kim, C., Seo, D.: Acoustic Emission Diagnosis System and Wireless Monitoring for Damage Assessment of Concrete Structures. In: NDT for Safety, Prague, Czech Republic (November 2007)

Deformation Detection in Structural Health Monitoring Pierantonio Merlino1 and Antonio Abramo2 1

PTLab - Agemont SpA, via J.Linussio 1, 33020 Amaro, Italy [email protected] 2 DIEGM - Universit` a degli Studi di Udine, via delle Scienze 208 - 33100 Udine, Italy [email protected] and ETH Lab - Eurotech Group, via Fratelli Solari 3/a - 33020 Amaro (UD), Italy [email protected]

Summary. In this contribution the issue of the embedded monitoring and detection of structural unhealthy conditions is addressed. The recent work on the design of a self-organizing architecture of sensing/communication nodes able to monitor the topological modification of structural surfaces is reviewed. The node design, purposely carried out to attain both contactless communication and sensing abilities, makes use of near-field coupling among nodes to implement both features, i.e. to monitor structure displacements and deploy a local communication network for the transfer of information. This technology shows low realization costs and lower power consumption compared to traditional wireless communication. A simple experimental setup is presented, demonstrating the architectural ability to trace the evolution of single structural fractures as well as of topological in-plane deformations, thus crediting the design as viable for the embedded surveillance of civil infrastructures.

1

Introduction

Complex civil infrastructures, such as bridges, buildings and dams, are often subject to severe environmental conditions and abnormal loads, e.g. strong winds, heavy rains, high humidity and huge temperature variations, that cannot be easily anticipated during their design [1]. This results into long-term structural deterioration that often is not detected by conventional visual inspection [2]. Moreover, catastrophic events, such as earthquakes, hurricanes or floods can severely affect the health of the structure and induce potential life-threatening conditions [3]. For this reasons, in recent years so called Structural Health Monitoring (SHM) technologies have emerged, opening interesting research fields inside the different branches of the engineering disciplines [1]. SHM systems involve large arrays of nodes that continuously monitor the quantities of interest by means of proper transducers, thus tracking the health of the particular structure. SHM systems are able to estimate the state of structural health and to evaluate the changes of the geometric properties [4, 5], S.C. Mukhopadhyay (Ed.): New Developments in Sensing Technology for SHM, LNEE 96, pp. 41–62. c Springer-Verlag Berlin Heidelberg 2011 springerlink.com

42

P. Merlino and A. Abramo

both in amplitude and on a time basis, that effect the performance of the monitored structure. Usual employment of SHM systems, indeed, are that of damage1 detection, damage localization and severity damage estimation [7]. Traditionally, SHM systems ground on wired-based technologies, but recently, the use of Wireless Sensor Networks (WSN) has emerged [7]. The reason stems from the large cost connected with the deployment of complex wired sensor arrays, which can be significantly reduced switching to their relatively cheaper wireless counterparts [1]. Moreover, wireless technologies are not subjected to wires wear and tear or damage caused by harsh weather conditions or other extreme events [3]. WSNs, in fact, do not require wire connections among sensors, and from them to the base-station, allowing the deployment of wide sensor networks even in almost inaccessible places. In addition, WSN can process the data collected by the sensors locally, and communicate summary information only [7]. These features — namely lower costs, cable-less installation and local computation capabilities — enable the deployment of hundreds of sensor nodes on a single structural element, thus enabling local-based damage-detection strategies [1]. A typical wireless sensor node architecture for SHM consists of a radio/ computing system to which specific transducers are attached, such as straingauges, accelerometers or others. As a consequence, the research on WSNs for SHM systems has mainly focused on new hardware architectures [8, 7], power consumption and node size reduction [8], or network organization [9] and implementation of distributed sensing and computation architectures and strategies [10]. Moreover, the recent development of very low power architectures for wireless sensor nodes has enabled the implementation of energy harvesting technologies [11] on SHMs systems.

2

Structural Deformation Detection Using Wireless Sensor Networks

A different approach for the solution of SHM problems was proposed in [12]. In this work a novel concept for the architecture of SHM nodes and network that exploits the near-field coupling between adjacent nodes for both the communication and the distance measurement was developed. The system is based on the idea that the deployment of a multitude of nodes on a structural element can enable the autonomous setup of a connected network of nodes that can, at the same time, monitor the local node-to-node displacements, hence acting as fracture monitor, and as a network on the whole, mapping the surface of the structural element under control on which the nodes have been laid, thus monitoring its surface deformations through the corresponding network topology modifications. 1

The term damage can be defined as changes introduced in the system that adversely affect its current or future performance. Its definition is commonly limited to changes to the material or geometric properties of the system [6].

Deformation Detection in Structural Health Monitoring

43

The single fractures can be detected and monitored locally through the measurement of the mutual distance between two adjacent nodes. Conversely, the monitoring of structural deformations requires the mapping of the structure geometry, hence the determination of all nodes relative positions. Consequently, as far as the structural deformation software procedure, the core of the system is the distributed localization algorithm described in [13]. Each monitoring node determines its position by means of the sole knowledge of the distance measured from its connected nodes, and based on the estimated positions those have made about their own positions. In this way each node can determine its positions in a relative coordinate system, from which the structure deformation can be monitored. A sketch of the described monitoring setup is shown in Fig. 1. The only requirement for the correct operation of the algorithm is the attribution to at least three of the nodes of the status of anchors, namely those nodes who know their exact (relative) position (e.g. the three unitary versors [1, 0, 0]T , [0, 1, 0]T , [0, 0, 1]T of what we called the relative coordinate system).

Fig. 1 A sketch of a near-field monitoring communication network

The algorithm of [13] assumes that each node collects the positions of the neighboring ones, temporarily acting as local anchors, and updates its own position through an asynchronous, decentralized optimization procedure. The localization algorithm guarantees the complete decentralization of the computation and implements a localization strategy largely insensitive to the peculiarities of the network topology.

44


As far as the hardware design is concerned, instead, the keystones of the proposed monitoring system are the near-field transmission technology adopted for the network communication, and the displacement measurement technique used for the estimation of the nodes relative distances. After the brief review of Sect. 3 on the near-field communication, clarifying some basic concepts about this technology, the problem of the antenna design and its design procedure are illustrated in Sect. 4. The distance measurement problem is analyzed in Sect. 5, where the adopted distance measurement technique based on Received Signal Strength (RSS) is also presented. The node architecture and its performance are presented in Sect. 6, which also describes the implementation of the network of monitoring nodes. Some recent improvements made on the measurement part of the analog circuitry are illustrated in Sect.7, before drawing the conclusions of the work in Sect. 8.

3

Near-Field Magnetic Communication

Since in SHM applications WSNs must guarantee operational lifetimes of several months, an important aspect to be carefully considered in designing the hardware architecture of the network nodes is their power dissipation. In WSNs, the largest part of power consumption is typically due to inter-node communication, especially for data/communication intensive applications [14]. For this reason we adopted near-field coupling between nodes as the underlying communication technology. To understand the reasons of this choice, it is worth to briefly introduce some basic concepts about near-field communication. The electromagnetic field associated with an antenna can be divided in two regions: the propagating far-field, and the non-propagating near-field [15]. The far-field, called Fraunhofer zone, is the region where the electromagnetic wave possesses a planar wave front, and where its electric and magnetic field are in phase. In this region the power of the radiation decreases as the square of the distance from the transmitting antenna, that is to say −20 dB/dec, and the absorption of the radiation power at the receiver has no effect on the transmitter itself. In the far-field the electromagnetic wave is fully formed and completely independent on the transmitting antenna [16]. Conversely in the near-field region, or Fresnel zone, the inductive and capacitive effects due to the currents and charges at the transmitting antenna are predominant[15]. In this case, the absorption of power at the receiver side has effects on the transmitting antenna also, so that the transmitter can even sense that power is absorbed from its emission. In near-field conditions the field strength decreases as 1/r3 , or −60 dB/dec, where r is the distance from the antenna. (See, e.g., Fig. 4.57 in [16]) Such a rapid roll-off of the field strength implies that inductive coupling effects are limited to a region relatively close to the antenna. As a consequence, wireless technologies based on near-field communication have a short range character, and typically do not interfere with other RF systems. The boundary between the near- and far-fields, that is to say the range of a near-field transmission, is determined by the kind and size of the antenna,


45

and by the frequency of the electromagnetic radiation. As a rule of thumb, for a coil antenna this boundary is set to be λ/2π, where λ is the wavelength of the electromagnetic field [16]. The short-range character of the near-field communication allows to attain significant reductions in the energy budget required for the communication among nodes, especially if non-radiative, resonant conditions are attained, thus increasing the operation lifetime that a WSN can feature [12].

4

Inductive Antenna Design

As presented in [12], we implemented a near-field transmission system using a PCB-based inductor antenna. A baseband communication is obtained between adjacent nodes by electrically modulating the magnetic flux of their coils with the transmitting data signal. Since the performance of the wireless communication, i.e. its communication range and datarate, depends mainly on the shape of the antenna and on its geometric dimensions, such as the track width and thickness, number and diameters of its turns, the accurate characterization and design are necessary in order to achieve the required figures. 4.1

Equivalent Circuit Analysis

The first activity of the whole design process was the modeling of the physical phenomena involved with near-field communication, and the first step of such a modeling activity was the identification of an equivalent circuit able to properly reproduce the inductive coupling between transmitting and receiving antennas. We adopted the circuit presented in [17, 18], which is depicted in Fig. 2. As can be seen, both the transmitting and receiving parts of the circuit are modeled as ideal inductors. However, the circuit is completed with series resistors representing the coils’ ohmic losses, and with a parallel capacitor that model the high-frequency shunt of the circuit. The

0 1 1 1 0 +0

VT 00 11 1 0 −

R T ILT

RR

+ CT

V LT LT

LR

00 11 00+ 0011 11

CR

VR 00 11 00− 011 1

− j ω MILT

Fig. 2 The equivalent circuit used for the modeling of the inductive coupling between adjacent coils

46


received voltage, VR, emerges from the √ presence of the current-controlled voltage generator, jωM ILT, where M =k LRLT is the mutual inductance between transmitting and receiving inductors, k is the coil coupling coefficient, and ILT is the current flowing through the transmitting coil. More complex circuits can be found in literature [19]. Nevertheless the adopted one, despite its simple structure, is able to describe the behavior of the near-field coupling with sufficient accuracy, and it can be effectively used to characterize the communication between adjacent nodes [12]. The analysis of the equivalent circuit, in particular of its step response, was performed to understand what the performance of the designed coil in terms of maximum datarate, power consumption, and communication range could be. Following [12], it can be found that2 : DR10% = RT / (2.3 LT) PTXmax = V02 / RT VRXmax = M VTXmax = k V0 LR / LT,

(1)

where DR10% is the maximum datarate3 , PTXmax is the peak transmitted power, VRXmax is the peak received voltage and V0 is the maximum input voltage level. The pulse duration can be reduced, hence the transmission datarate increased, using coils characterized by low inductance and high resistance values. Moreover, the maximum datarate can be finely tuned adding a resistor RS of proper value in series with the coil. Since CT and CR are low enough to be neglected, the resistance RS results in series with the equivalent resistance of the coil, RT, allowing the trimming of the datarate. Unfortunately, high values of the tuning resistor RS tend to reduce the inductor peak voltage, VTXmax, thus decreasing the effective communication range. In fact, if a voltage step with a rising time of tr is applied to the input, that is to say: vT(t) =

V0 t × (1(t) − 1(t − tr )) + V01(t − tr ), tr

(2)

where 1(t) is the step function, then the voltage emerging at the inductor, VLT, is going to be: vLT =

V0 LT (1 − e−t/τ )1(t) − tr R T + R S V0 LT − (1 − e−(t−tr )/τ )1(t − tr ) + tr R T + S S + V0 e−(t−tr )/τ 1(t − tr )

2

3

(3)

The formulas of [12] are computed under the assumption of low CT and CR values (≤ 10 pF). Since the voltage VT at the coil terminals is a train of bipolar pulse signals [12] whose positive parts correspond to the rising edges of the input and whose negative ones correspond to its falling edges, the datarate is computed as the inverse of the pulse time length at 10% of the pulse peak voltage, DR10% = 1/t10% . Assuming that vLT(t10% ) = vT0 e−t10% RT/ LT = 0.1 V0, it can be found that the pulse time length is t10% = 2.3 LT/ RT.


47

where τ = LT/(RT + RS). As can be noticed, the maximum VLT of the transmitting antenna is obtained when t = tr and VLTmax = Vtr0 R L+TR (1 − e−t/τ ). T S Consequently, from Eqs. (1)-(3) it can be observed that if high RS values are chosen, the datarate increases but the communicationrange decreases due to the reduced received voltage VRXmax = M VTXmax < k V0 LR / LT. 4.2

Antenna Modeling and Design Strategy

Having analyzed the equivalent circuit of the near-field coupling, the design of the transmitting and receiving antenna can be performed so as to obtain the required performances in terms of datarate, power consumption and communication range. Using the simulator presented in [12], we are able to extract the electrical parameters of planar, multi-turns, rectangular and circular coil configurations. For the sake of clarity, we briefly review here some results about coil dimensioning. For a coil of arbitrary shape realized with the PCB technology, the geometrical parameters used for the design are the trace width, w, and thickness, h, of the deposited metal, the number of turns of the inductor, n, and the diameter(side) of the circular(square) coil, d. Fig. 3 shows how the datarate depends on the geometrical parameters of the coil. As expected, the datarate increases at decreasing values of both w and h, since this corresponds to high values of the series resistance RT. However, this condition results into the reduction of the magnetic coupling of the coils. (See Fig. 4-(c),(d)). It must be also noticed that the coupling coefficient k, hence the received voltage VR, depends additionally on the coil diameter, d, and on the number of turns, n. (See Fig. 4-(a),(b)) Since high values of k are desirable, wide diameters and high number of turns should also be designed. As a consequence, a trade off between the geometric parameters of the inductor is necessary. Based on the considerations above, we implemented a simple procedure that, accounting of the performance requirements and the geometrical limits 8

2 (a)

(b) 6

0 4

3 2

2

Data Rate [Mbit/s]

Data Rate [Mbit/s]

1

1 0

0.2

0.8 0.6 0.4 Trace Width [mm]

1

20

80 60 40 Trace Thickness [ μm]

0 100

Fig. 3 The communication datarate plotted as a function of the coil geometrical parameters.

48

P. Merlino and A. Abramo 0.05

0.08 0.06

0.03

k

k

0.04

0.04

0.02 (a) 0.01

2

4 6 8 Diameter [cm]

10

(b) 2

4

6 Turns

0.02 10

8

0.04 0.036

0.036

k

k

0.038

0.035

0.034 (c) 0.032

0.2 0.4 0.6 0.8 Trace Width [mm]

(d) 0.034 80 100 60 40 20 Trace Thickness [ μm]

Fig. 4 Coupling coefficient, M, between two identical coils, plotted at varying coil diameter and number of turns.

of the PCB technology, searches the optimal values of the inductor geometrical parameters. The strategy, depicted in Fig. 5, can be described as follows: Step 1: Given a specific application, the minimum datarate, DRmin, the minimum received voltage4 , VRXmin, and the maximum power consumption, PTXmax, are determined. The technology chosen for the antenna realization determines the span of feasible values of the geometrical parameters. In our application we chose the PCB technology and set feasible values for the diameter, Id, trace width, Iw, thickness, Ih, and number of turns, In. Step 2: The actual peak received voltage, VRX, is computed setting the geometrical parameter at their minimum values. It must be noticed that we decided to implement circular inductors for both the transmitting and receiving coils in order to obtain isotropic field conditions around the circuits. This situation is favorable for the nodes reciprocal distance, as will be seen in Sect. 6. Finally, to reduce power consumption, the tuning resistor, RS, is initially set to its maximum value, RSmax. Step 3: If the peak received voltage is larger than its allowed minimum, that is to say VRX ≥ VRXmin, the procedure proceeds with the next step; as a matter of fact, this condition ensures that the communication range requirements are satisfied. Otherwise, the tuning resistor is reduced by a fixed small decrement, ΔRS, and the procedure iterates until the condition is met. 4

The minimum received voltage is computed given a desired signal-to-noise-ratio and communication range, as will be explained in the next Subsection.


49

Fig. 5 Flowchart of the antenna design procedure.

Step 4: The actual datarate and power consumption are computed. If their values meet the requirements, that is to say DR ≥ DRmin and PTX ≤ PTXmax, then the antenna geometric parameters are identified and the procedure stops. Otherwise, Step 2 is resumed with incremented values of the geometric parameters. Since compact antennas are desirable, the procedure was set to privilege small antenna designs first. Fig. 6 shows the results obtained with the procedure in our case, i.e. the dependence of the geometrical parameters of the circular coils on the communication range. As expected, both diameter and turns must increase at increasing communication range, since the mutual inductance must be increased with distance. In our case the design strategy found the optimal geometric parameters as d = 50 mm, n = 8, w = 0.2 mm, h = 50 μm. The circular inductors overlapped on 4 PCB layers. The tuning resistance was set at 3kΩ, yielding a maximum datarate of 20 Mbit/s. The maximum communication range was found to be 10 cm, while the peak power consumption for the transmission was estimated in 8 mW.


Diameter d [mm]

50 50 40 30 20 10

Turns n

8 6 4 2 0

1

2

3

4 5 6 Distance [cm]

7

8

9

10

Fig. 6 Behavior of the geometric parameters as a function of the communication range.

4.3

Transmission Characterization

Since the performance of a digital communication is typically defined in terms of bit-error-rate (BER) and signal-to-noise-ratio (SNR), it is important to understand the effect of the geometrical size of the inductors on these performance figures. However, the equivalent circuit of Fig. 2 is not devised for the estimation of SNR and BER. Consequently we developed an additional model for the simulation of the whole transmission system [12]. The model was implemented in MATLABTM, and was describing the communication chain as the cascade of the filter elements shown in Fig. 7.

bk

VT

Z

R

TX

ILT

+

ILT ε

ε ^ b k Z

VR

RX

R

Fig. 7 Schematic view of the near-field communication system as modeled in software.

The first stage is an interpolation filter, generating the transmitted voltage, VT, from the input bit stream, bk. VT is then elaborated by the transmitting block, TX, whose transfer function: ILT 1 = VT RT + jωLT

(4)


51

provides the transmission inductor current, ILT. Due to the short range of the transmission we decided to model the channel simply as an additive white Gaussian noise, ε. Proceeding along the communication chain, the receiving stage, RX, is responsible for the conversion between the received current, ILTε, and the corresponding received voltage, VR, which is obtained through the following transfer function: VR jωM = . (5) ILTε −ω 2 LRCR + jωRRCR + 1 Finally, the received voltage VR is sampled and the received bits are decided using a threshold transfer function. Fig. 8 shows the simulation results obtained in case of circular coils of different diameters. As can be seen, the SNR (thus BER) values vary accordingly with coil diameters: wider coils guarantee higher SNR values thanks to stronger coil coupling.

35

30 d = 75mm

25

SNR [dB]

−6

BER=10

20 d = 50mm 15 −5.4

BER=10 10

5

d = 25mm −1

BER=10 0

0

2

4 6 Distanza [cm]

8

10

Fig. 8 SNR vs. communication distance for circular coils of different diameters, d

The BER values can be used to define the maximum communication range. For example, targeting a desired BER of 10−1 , the communication range exceeds 10 cm for a 50 mm circular inductor if a SNR = 0 dB is targeted.

5

Distance Measurement

In our application, the main ingredient for attain the desired network structural monitoring is the measurement of nodes relative distances. To this purpose, it is worth summarizing some of the most common ranging techniques that are used in WSNs before presenting the implementation of the one we chose for our case-study.

52

5.1


Distance Ranging Techniques in WSNs

Ranging techniques can be divided into several classes, depending on the physical principles exploited to measure the distance between nodes [20]. The more used techniques are based on the propagation time of the electromagnetic wave, or on the received power. The first class includes Time Of Arrival (TOA), Time Difference Of Arrival (TDOA) and Round-Trip time Of Arrival (RTOF), while the RSS is included in the second class. Time-based Techniques Time based techniques exploit the propagation time of the radio signal from the transmitter to the receiver nodes. Since the propagation speed of an electromagnetic wave in the medium is c, the distance can be simply computed as d = c × t, where d is the distance and t is the time of propagation. Therefore, the measure of distance between nodes can be obtained by measuring the propagation time of the transmitted signal. The simplest way to measure distance in WSNs is the TOA technique. The measurement setup is made of a transmitter node, sending a data packet that incorporates the sending timestamp, and a receiver one that computes the propagation time of the signal as the difference between the received timestamp and the one obtained at the packet arrival. This technique provides an accuracy of microseconds, but it requires the time synchronization between the sender and receiver. In order to bypass this requirement, the RTOF technique can be adopted. In this case, the transmitter sends a wave pulse and waits for the wave reflected by the target. The propagation time is computed by halving the wave roundtrip time. Alternatively, the receiving node can act as a transponder, sending back a reply wave upon reception. This variant guarantees the presence of stronger returning signals, but the delay inherent to the resending process can introduce errors in the time determination. The last time-based technique, TDOA, uses a minimum of three nodes. A transmitter sends a signal that is received by two receiver nodes, and the propagation time is computed using the time arrival differences. The TDOA algorithm assumes that the nodes are time synchronized and that they know their locations. Received Signal Strength Technique The RSS measurement method is based on the dependence of the field strength on the distance from the transmitting antenna. The measure of the distance is performed relating the signal strength at the receiver with a model of the decay over distance of the field strength. A commonly adopted evaluation of the decay is based on the path loss model (PL) [21]: PL(d) = 10 log10

Pt d 4πd0 = 10 N log10 + 20 log10 + χσ Pr d0 λ

(6)

where Pt and Pr are the transmitted and received power, respectively, d is the distance, d0 is a reference distance (usually d0 = 1 m) and λ is the wavelength


53

of the field, N is the path-loss coefficient that takes into account the specific propagation environment and others effects such as fading or multipath, and finally χσ is a zero-mean Gaussian process with standard deviation σ, modeling measurement noise. Since this technique does not require any synchronization between nodes, its implementation is simple and it is widely adopted in WSNs. However, the RSS technique suffers from multipath inaccuracy especially in indoor environments, and the PL coefficient can greatly vary making this range technique very inaccurate. Nevertheless, we judged that in our case the short-range character of the envisioned application could strongly mitigate both the multipath and distortion effects, and that the measurement accuracy could be more than satisfactory for our application. For this reason, we decided to choose the RSS in consideration of its simpler structure. As a final remark, it can be noticed that Eq. (6) is valid only in far-field conditions and cannot be used in the present near-field case. For this reason in the next Section we will make use of the Eqs. (1) and of the simulation framework previously described for the characterization of the measurement technique. 5.2

RSS Ranging Technique Implementation

360

Mutual Inductance Measured VR

320

Simulated VR

7.5

280 6.5

240 200

4.5

160 120

2.5

80

Mutual Inductance [ μH]

R

Received Voltage V [mV]

As previously mentioned, Eqs. (1) show that the received voltage, VR, is proportional to the mutual inductance, M, and how the latter strongly depends on the distance between transmitting and receiving coils. Consequently, distance information can be extracted from the transduction of M values, i.e. from the measurement of the received signal level. To this purpose, we used the simulation framework previously described to quantify the dependence of the mutual inductance of nodes on their mutual distance.

40 5

7

9

Distance [cm]

11

13

0.05

Fig. 9 Measured (circles) and simulated (line) received voltage as a function of the distance between circular coils. The coil parameters are: d = 50 mm, n = 8, w = 0.2 mm, h = 50 μm. The plot of the mutual inductance is also included.

54


As a first step we tabulated the VR vs. r (i.e. M vs. r) relationship, as experimentally extracted using two template inductors. A look-up table was then stored on the node, and used to convert M variations, measured at run-time, into the correspondent distance measure. (See Fig. 9 for the case of a circular coil). Since the mutual inductance of adjacent coils, and consequently the received voltage level, do not vary linearly with distance, but rather as 1 / r3 (see Fig. 9), the voltage variations ΔVR corresponding to equal distance variation Δr is larger at small r, that is to say when the coils are closer. This evidence sets the limits to the precision of the measurement, since the system is able to resolve small distance variations only at close distance, where the mutual inductance is high. The interested reader can found a more detailed explanation of this point in [12].

6

Node Design and Network Implementation

In order to validate the proposed near-field sensing/communication strategy, a few versions of the node were realized [12]. (See Fig. 10 showing the second circular prototype. For simplicity, a first version had been realized in square shape, but it was soon abandoned because of the limited communication range and the anisotropic character of the electromagnetic field distribution.) The

Fig. 10 Circular prototyping board for the evaluation of the ing/communication monitoring system: 8-turns external, concentric w = 0.2 mm; h = 50 μm; d = 5 cm

senscoils;


55

system showed encouraging results both in terms of power consumption and datarate, as well in terms of accuracy of the distance measurement. The node was designed following the indications of the simulation framework described above, on whose prediction ability we gained high confidence thanks to the very good comparison between simulations and experimental verification. The digital part of the design is based on the use of a NXPTM LPC2106 microcontroller, where the main functions of the systems, such as distance computation, positioning algorithms and communication through nodes and to a host PC, are implemented in software. The choice of the specific microcontroller was made considering its computational capability, the availability of several on-chip peripherals (such as counters, UART and SPI ports, which were found very useful for PC and ADC interfacing, actually allowing radical simplifications on the PCB layout), and its limited power consumption, obtained thanks to the presence of an efficient power management unit. As far as the analog part of the circuit5 , the near-field transmitter is based on a simple power-CMOS inverter directly driving the transmission coil, whose impedance was adapted by means of a tuning resistor, RS. (See Fig. 11) The receiver, instead, was set up in a two-stage topology. (See Fig. 12). The first stage acts as a differential amplifier, boosting the received differential voltage, VRX, while the second one is a Schmitt-trigger, where the positive/negative received pulses cause the positive/negative saturation of the amplifier and reshape the received signal into a square pulse train. As a result, the output of the second stage is a NRZ signal identical to the transmitted data, but with a voltage level that relates, as said, to nodes mutual distance. The characteristics of the near-field communication were experimentally tested, yielding6 a datarate of 20 Mbit/s. In very good agreement with the performed simulations, we found that the coupling coefficient started from a value of 0.13 for nodes that were positioned so as to touch, and decreased down to 0.01 for a node distance of 10 cm. Consequently, a maximum communication range of 10 cm was set. (See Fig. 8). In order to convert RSS measurements into distance, the peak voltage of the received digital signal, i.e. of the output of the Schmitt-trigger, is extracted by the simple diode-capacitor peak detector. (See again Fig. 12) This maximum value is then converted by the microcontroller using its internal ADC, and from this value the voltage-to-distance correspondence is extracted using the lookup table that, as said, was initialized with the mapping of the VR vs. distance characterization. In terms of resolution, the measurement circuit was able to 5

6

All transceiver circuits were simulated in SPICE to evaluate their functionality and performances. The magnetic coupling was modeled using a conventional transformer, whose leakage inductance was set to a value greater than the magnetizing one to account for the rather low values of the coupling coefficient and of the mutual inductance that are obtained in the actual inductors. This result was obtained with a tuning resistor RS = 3kΩ.

56


VDD

RS

Data TX

R +

+

VTX

R

+

+

−

V RX −

−

Data RX

R

R

−

R R

Fig. 11 Schematic view of the transmitting part of the transceiver circuit

R + V RX −

R + +

R

D

− −

C

R Fig. 12 The distance measurement circuit based on a peak detector

discriminate distance variations with an accuracy that, being dependent on node distance, was found to range between 48 μm at zero distance and 480 μm at 10 cm. Since communication among nodes is necessary in the case of the monitoring of isolated fractures as well as in that of surface mapping, a protocol able to resolve the contention originating from the possible simultaneous node access to the communication channel was included, so as to properly treat the second case [16]. The simultaneous access to the channel, indeed, can hamper the nodes communication and irreparably corrupt the distance measurement. Since the real-time requirements of the considered applications are rather limited, we decided to favor the simplicity of the software implementation at the expenses of the execution performance. Consequently, we chose to implement a simple token-based communication protocol to resolve the medium access contentions: only the node who owns he token can start broadcasting the data to the adjacent ones, who can only reply to an external request. Once the communication is completed, the token is passed to one of the adjacent node. If the token is repeatedly exchanged in a proper sequence, all nodes of the network will eventually receive it, and will accomplish to their software tasks. To guarantee a democratic periodicity it is sufficient to annotate each node with a unique network identifier. To this purpose we developed a fully decentralized identifier assignment procedure, specifically derived for the case


57

(a)

(b) Fig. 13 (a): Demonstrator of the SHM system. (b): Close-up of the graphical sketch on the PC: the crack at the up-right corner of the marble tile widens proportionally as nodes are manually displaced in reality.

58


of token-based communication protocols (see [12] for more details): at reset the network proceeds to the unique node identification; then each nodes discovers its first neighbors and stores their identification numbers; the token is then introduced into the network, and based on its possession the nodes start exchanging the estimates of their relative positions, measured locally through RSS conversion; the procedure endlessly iterates, so as to converge to the actual mutual distances of the nodes [13], and tracking any topology modification that may occur to the network. As a prototype demonstrator of a surface mapping application, the basic set up shown in Fig. 13(a) was arranged: a set of three nodes is mimicking the mapping of a marble tile surface, where an existing crack has to be monitored; the mutual distances are computed by the nodes following the procedure sketched above, and the crack dimension is displayed on the GUI of a laptop PC connected to one of the node via USB connection (see Fig. 13(b)); the network continuously updates nodes positions, and so does the GUI as far as the plot of the crack gap is concerned; when some of the nodes are manually displaced, the GUI enlarges or reduces the crack dimension according to the real node displacement.

7

RSS Circuit Improvements

After several test session, it was found that the design of the analog part of the circuit was hampered by severe limitations, mainly originating from the simple peak detector implementation. In fact, the pinning of VR to its maximum value, that in [12] was obtained through the use of a simple diode-based peak detector, was shown to limit the dynamic range of the measurement, besides introducing a severe sensitivity of the very same measurement on temperature, attributed to the drift of the diode parameters. Although alternative circuit topology exist that can limit this detrimental effect [22], we decided to switch to the design of a RF detector, due to its higher dynamic range (> 40dB) and good temperature stability. Consequently, to improve the performance of the measurement circuit we implemented the circuit of Fig. 14 that, as can be seen, makes use of the cascade of a Variable Gain Amplifier (VGA) and of a RF power detector. C +

C C

+

V RX

RF detector C

VGA C

−

−

+ R

C C

C

−

Vgain R

R

Fig. 14 The distance measurement circuit

Vmeas C filter


59

In a VGA the gain is set by an external control voltage. In Fig. 14, a simple resistive trimmer is introduced to this purpose, but in reality we made use of the ADC available on the microcontroller to set the control voltage of the VGA. In this way, we were able to properly adjust the received voltage levels to the input requirements of the RF detector, and to compensate the errors introduced by PCB manufacturing mismatches. The core of the RSS circuit is the RF power detector, extracting the power of its RF input. Its output, indeed, returns a voltage that is proportional to the power of the incoming signal that, in our case, corresponds to the RSS, which can be directly used to extract distance measures. As can be seen in Fig. 14, a filter capacitor was placed at the output of the RF detector, so as to limit the output voltage ripple to provide a more stable signal to the ADC, and to filter the high frequency noise that can be spuriously collected at the receiving antenna and amplified by the VGA. To demonstrate the validity of the design we implemented it using commercial parts, assembled in the discrete test setup shown in Fig. 15. Here, the circular nodes previously developed were used only as antennas, that is to say completely bypassing their circuitry. As can be seen, the transmitting antenna is driven by a square wave, provided by an arbitrary signal generator (not shown), simulating the data transmission. The receiver coil is connected to a VGA board, and this, in turn, to the RF detector by means of coax connection. An impedance matching circuit was also implemented on the RF detector in order to avoid wave reflections. Finally the output voltage, i.e. the RSS, is extracted and displayed at the oscilloscope (not shown).

Fig. 15 RSS system setup used for evaluation and testing purposes

60


Table 1 Power dissipation figures of the prototyping system. On the left: Lifetimes of the different system components at the estimated operation duty=cycles. A battery featuring 1 Ah was assumed. On the right: instantaneous power dissipation of the different system components.

Duty Cycle Lifetime [h] 100% 50% 4% 1%

48.3 96.6 1207.5 4830

Node Power functionality Consumption [mW] CPU 18 Transmission 8 Reception 27 Measurement 86.7

As far as the VGA component is concerned, we chose the Analog DevicesTM AD8330 because of its wide input dynamic range, low power and low noise characteristics. The chosen RF power detector, instead, was the Analog DevicesTM AD8310, featuring high dynamic and frequency ranges, and large temperature stability. The measurement circuit of Fig. 15 was then cascaded to a TITM ADS7884, that is a 10-bit, 3 MSamples/s ADC that we introduced for analog-to-digital conversion, and whose data samples are sent to the microcontroller for RSS vs. distance conversion. The whole monitoring system was designed taking power dissipation in particular consideration. (See Tab. 1) The near-field transmission part of the analog circuitry can be considered rather parsimonious, requiring just 8 mW (1.6 mA@5 V) to operate, sensibly less than what required by typical ZigBeeTM transmitters. The receiver circuit power dissipation, mainly due to the two operational amplifiers, amounts to 27 mW. The microcontroller, instead dissipates a peak value of 18 mW (10 [email protected] V). The largest contribution to power dissipation, however, is to be ascribed to the distance measurement circuit, and amounts to 86.7 mW, partitioned as follows: 52.8 mW for the VGA operation, 26.4 mW for the RF power detection, and 7.5 mW for the ADC. However, since the RSS circuit can operate with very small duty-cycles (< 1%), as it is required by the localization algorithm, this rather large power dissipation does not significantly influence the overall lifetime of the system. This can be seen looking at the second column of Tab. 1, where lifetimes were computed assuming a 1 Ah battery, whose dimensions are compatible with the integration onto the PCB. For the sake of completeness, it must be noticed that the severe constraint imposed to lifetime by the microcontroller, limiting to about two days the operation of the system, can be be radically improved by aggressive power saving strategies. However, even a simple reduction of the microcontroller duty-cycle, assumed here to operate at full throttle for worst-case considerations, can significantly increase the nodes’ lifetime to values more compatible with long-term applications.


8

61

Conclusions

To summarize, this contribution demonstrates that embedded monitoring of structural elements can be performed using low-complexity, low-cost, and low-dissipation systems. In other words, that SHM requirements are compatible with WSN specifications. Although applied to a laboratory case-study, the results obtained envision the deployment of large networks of monitoring nodes that, self-assembling as a network on the whole, can constantly trace the displacements of structural parts, or even map their topological deformations. Acknowledgments. The work was partially funded by ETH Lab, Eurotech Group, Amaro (UD), Italy.

References [1] Lynch, J., Loh, K.: A summary review of wireless sensors and sensor networks for structural health monitoring. In: Shock and Vibration Digest, vol. 91, Sage Publications, Thousand Oaks (2007) [2] Sazonov, E., Janoyan, K., Jha, R.: Wireless intelligent sensor network for autonomous structural health monitoring. In: Smart Structures and Materials 2004: Smart Sensor Technology and Measurement Systems, vol. 5384, p. 305 (2004) [3] Bocca, M., Cosar, E., Salminen, J., Eriksson, L.: A reconfigurable wireless sensor network for structural health monitoring. In: Meier, U., Havranek, B., Motavalli, M. (eds.) Proceedings of the 4th International Conference on Structural Health Monitoring of Intelligent Infrastructure, Zurich, Switzerland. International Society for Structural Health Monitoring of Intelligent Infrastructure (2009) [4] Kim, S., Pakzad, S., Culler, D., Demmel, J., Fenves, G., Glaser, S., Turon, M.: Health monitoring of civil infrastructures using wireless sensor networks. In: IPSN 2007: Proceedings of the 6th International Conference on Information Processing in Sensor Networks, pp. 254–263. ACM Press, New York (2007) [5] Rizzoli, V., Costanzo, A., Montanari, E., Benedetti, A.: A new wireless displacement sensor based on reverse design of microwave and millimeter-wave antenna array. IEEE Sensors Journal 9, 1557–1566 (2009) [6] Farrar, C., Worden, K.: An introduction to structural health monitoring. Phil. Trans. R. Soc. A 365, 303 (2007) [7] Wu, J., Yuan, S., Zhao, X., Yin, Y., Ye, W.: A wireless sensor network node designed for exploring a structural health monitoring application. Smart Material and Structures 18, 1898 (2007) [8] Liu, L., Yuan, F.: Wireless sensors with dual-controller architecture for active diagnosis in structural health monitoring. Smart Material and Structures 17, 25016 (2008) [9] Kottapalli, V., Kiremidjian, A., Lynch, J., Carryer, E., Kenny, T., Law, K., Lei, Y.: Two-tiered wireless sensor network architecture for structural health monitoring. In: Proceedings SPIE, vol. 5057, p. 8 (2003)

62


[10] Mitchell, K., Sana, S., Liu, P., Cingirikonda, K., Rao, V., Pottinger, H.: Distributed computing and sensing for structural health monitoring systems. In: Proceedings SPIE, vol. 3990, p. 156 (2000) [11] Park, G., Rosing, T., Todd, M., Farrar, C., Hodgkiss, W.: Energy harvesting for structural health monitoring sensor networks. Journal of Infrastructure Systems 14, 64–79 (2008) [12] Merlino, P., Abramo, A.: An integrated sensing/communication architecture for structural health monitoring. IEEE Sensors Journal 9, 1397–1404 (2009) [13] Abramo, A., Blanchini, F., Geretti, L., Savorgnan, C.: A mixed convex/nonconvex distributed localization approach for the deployment of indoor positioning services. IEEE Transaction on Mobile Computing 7, 1325–1337 (2008) [14] Enx, C., Scolari, N., Yodprasit, U.: Ultra low-power radio design for wireless sensor networks. In: Proceedings RFIT, vol. 1 (2005) [15] Evans-Pughe, C.: Close encounters of the magnetic kind (near field communications). IEE Review 51, 38–42 (2005) [16] Finkenzeller, K.: RFID handbook: fundamentals and applications in contactless smart cards and identification. Wiley, Chichester (2003) [17] Miura, N., Mizoguchi, D., Sakurai, T., Kuroda, T.: Analysis and design of inductive coupling and transceiver circuit for inductive inter-chip wireless superconnect. IEEE Journal of Solid-State Circuits 40, 829 (2005) [18] Miura, N., Mizoguchi, D., Inoue, M., Sakurai, T., Kuroda, T.: A 195-Gb/s 1.2-W inductive inter-chip wireless superconnect with transmit power control scheme for 3D-stacked system in a package. IEEE Journal of Solid-State Circuits 41, 23 (2006) [19] Neagu, C., Jansen, H., Smith, A., Gardeniers, J., Elwenspoek, M.: Characterization of a planar microcoil for implantable microsystems. Sensors and Actuators A: Physical 62, 599 (1997) [20] Diao, Y., Fu, M., Zhang, H.: An overview of range detection techniques for wireless sensor networks. In: 8th World Congress on Intelligent Control and Automation (WCICA), pp. 1150–1155 (2010) [21] Rappaport, T.: Wireless communications: principles and practice, 2nd edn. Prentice-Hall, Englewood Cliffs (2001) [22] Rixon, A., Waugh, R.: A suppressed harmonic power detector for dual band phones. Applied Microvawe and Wireless 11 (1999)

MEMS Strain Sensors for Intelligent Structural Systems Debbie G. Senesky and Babak Jamshidi University of California, Berkeley

Abstract. The use of microelectromechanical systems (MEMS) technology to develop strain sensors (resonant and capacitive) is the main topic of this paper. Sensing technology can advance the design and integrity of structural systems in various industries by enabling monitoring of strains and stress concentrations within a mechanical structure in real-time. MEMS-based strain sensors enable performance improvements through increased resolutions, increased operation bandwidths and reduced sensitivity to noise. Therefore, the application of these devices can significantly improve the design robustness and efficiency by predicting catastrophic failures and enabling lightweight designs. MEMS strain sensors can impact the oil and gas, automotive, aerospace and buildings industries through the real-time monitoring of critical components. In addition to device performance, packaging, temperature compensation and long-term drift are important design considerations. Keywords: sensors, MEMS, strain sensing, stress, structural monitoring.

1 Introduction Since the early 1990’s, the use of semiconductor fabrication techniques to create microelectromechanical Systems (MEMS) has been explored to create robust devices with decreased footprints at reduced costs [1]. Recently, the use of MEMS sensing devices in wireless sensor networks (WSN) has been explored to collect data from structures and environments in real-time [2,3]. One vision of the future is to utilize sensing technology to improve the design and efficiency of mechanical structures through real-time, condition-based monitoring [4]. Delamination, mechanical fractures and corrosion can be identified with proper implementation of sensing systems. This paper will review the design and fabrication of MEMS strain sensors that can be utilized to create intelligent structures. MEMS-based designs are advantageous due to the ability to create devices with high resolution and low noise. In addition, the high-volume manufacturability of MEMS devices reduces the production costs to a fraction of existing monitoring systems. Strain sensors with resonant and capacitive sensing schemes are typically employed and will be described here. It is important for the designer to consider the entire system and to choose a sensing method which exhibits the least environmental S.C. Mukhopadhyay (Ed.): New Developments in Sensing Technology for SHM,LNEE 96, pp. 63–74. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

64

D.G. Senesky and B. Jamshidi

biasing (e.g. temperature). Advancements in packaging and electronics are required to enable monitoring systems that are robust and calibrated for offset drifts. It should be noted that metal-foil strain gauges are commercially available for structural monitoring. Metal-foil gauges utilize the piezoresistive effect to transduce or convert a mechanical strain to an electrical signal. These devices typically have a large footprint and low resolution. Also, piezoelectric devices (e.g. surface acoustic wave or bulk acoustic wave) can be used to monitor changes in the acoustic emission of a material to observe mechanical degradation. This paper focuses on MEMS strain gauges that utilize free-standing, released sensing elements for direct strain detection and strain rate monitoring.

2 Applications for MEMS Strain Sensors MEMS sensors are advantageous due to the ability to obtain extremely small sensor footprints and batch fabricate complex structures. Strain sensors enable monitoring of strains and stress concentrations and can be utilized for condition-based monitoring of structures. More specifically, these devices can be used to predict and assess mechanical failures such as delamination, mechanical fatigue cracks and corrosion. Strain sensors can be applied to improve the performance of systems used in various industries such as • Automotive • Aerospace • Buildings and bridges • Machining tool • Wind turbine power. The improvement of these systems can have a significant impact on society by improving safety, increasing operation lifetimes and enabling energy efficiency. For example, strain sensors can be embedded into composite materials to monitor degradation and prevent catastrophic failures [5-6]. In addition, strain monitoring can be used to develop efficient and cost effective designs with decreased safety factors due to more accurate and reliable assessment of structural health. However, it is critical to carefully integrate the strain sensor and circuitry into the system to prevent undesirable and unpredictable failures [7]. It should be noted that the packaging and positioning of strain sensors should be engineered to enable strain transfer to the sensing element without impact to the performance of the underlying structure. Stress concentrations and delamination of existing structures should be avoided. Therefore, loading conditions should be determined for accurate measurements. In Fig. 1 schematic images of positioning strain sensors onto mechanical structures to monitor axial and torsional loading are detailed. Clearly, alignment to the structure is critical and should be considered. In addition, in-plane behaviours should be studied to ensure appropriate strain transfer to the sensing element. Therefore, packaging and bonding technology should be developed to enable efficient integration with the overall system.

MEMS Strain Sensors for Intelligent Structural Systems

65

Fig. 1 Schematic images of configuring strain sensors on a mechanical structure to measure axial (a) and torsional loads (b)

3 Fabrication Technologies The fabrication of MEMS sensors leverages processes developed by the semiconductor industry for integrated circuits (ICs). 2-dimentional patterns along with additive and subtractive processes are used to define 3-dimentional structures. Fig. 2 depicts a simplified fabrication sequence that can be used to design MEMS sensors. Typically, a silicon (Si) substrate is used as the base material for the device. It can also serve as electrical interconnect to the MEMS structures if the Si substrate is highly doped. A sacrificial material such as silicon dioxide (SiO2) or a polymer thin film (e.g. photoresist or SU-8) is used as a temporary material to enable growth of the structural layer. To enable access of the structural material to the underlying substrate the sacrificial material is etched or patterned. The structural layer (polycrystalline Si in our example) is deposited and subsequently etched to define the top view geometry of the device and to enable access to the underlying sacrificial material. A dry release technique is preferred to avoid stiction effects. Therefore, vapour (dry) release with hydrofluoric (HF) vapour or XeF2 is often utilized to etched the sacrificial material. To create electrical interconnect to the MEMS device, thin film metals are deposited (sputter deposition or evaporation) onto the structure. It should be noted that this fabrication sequence can be modified to implement new materials to enable alternative sense mechanisms. The packaging of the device can also use semiconductor fabrication techniques. Designers that are interested in fabricating devices but do not have access to a clean room facility can utilize a multi-user fabrication processes such as MEMSCAP’s polysilicon multi-user MEMS processes (polyMUMPS) [8].

66


Fig. 2 Illustration of a simple fabrication process used to create released MEMS structures. It should be noted that packaging and electrical interconnect are not shown in this schematic

Furthermore, new materials can be implemented in the design of MEMS strain sensors to extend the performance of devices to extreme temperatures. For example, silicon carbide (SiC) is being utilized in the design of strain sensors [9-10]. In addition, advanced piezoelectric materials (e.g. PZT or AlN) have been extensively studied to design sensors that can monitor acoustic emissions [11].

4 Resonant MEMS Strain Sensors MEMS strain sensors utilize free-standing mechanical elements to detect physical changes or deformations within interfacing structures. Observing shifts in the resonance frequency of a MEMS structure is one method for detecting strain and often leads to high sensitivity measurements with increased bandwidths and lownoise [12]. The shift in the resonance frequency with applied strain (sensitivity) can be measured and monitored with external electronics (e.g. oscillator circuits) [13-14]. A double-ended-tuning-fork (DETF) structure can be fabricated using MEMS processes and designed such that the resonant frequency of the DETF tines is highly sensitive to external forces. In such a design scenario, the DETF is used as the sensing element for strain detection. The DETF can be driven into resonance with electrostatic forces via interdigitated comb drives (Fig. 3) or parallel capacitive plates. A secondary set of electrostatic features are used for sensing changes in frequency. Fig. 3 details the design features of a MEMS strain sensor with interdigitated drive and sense electrodes and a DETF sensing element. The DETF structure can be modelled as two clamped-clamped beams. The axial force applied to a single clamped-clamped beam is proportional to the strain (ε),


67

Fig. 3 SEM image of MEMS resonant strain sensor with a DETF sensing structure ad interdigitated comb features used for drive and sense

2

(1)

where w is the beam width, t is the beam thickness and E is the Young’s modulus of elasticity. Recalling from basic structural mechanics, is the ratio of the change in length and the original length, ∆

(2)

where is original length of the beam and is the final length of the beam (Fig. 4) [15]. Although changes in length are shown in Fig. 4, it should be noted that out-of-plane deformations contribute to the behaviour of the mechanical structure.

Fig. 4 Schematic image of undeformed DETF structure and DETF structure subjected to axial loading causing changes in the device dimensions

68


Ultimately, the deformation in the DETF structure from an applied force causes a shift in the resonant frequency. The resonant frequency, fr, of a clamped-clamped beam is described by (3) is the effective mass of the beam and is the effective spring conwhere stant of the beam. It should be noted that the built-in strain ( ) and the mass of ) which in this analysis are assumed to be lumped the drive-sense actuators ( at the center of the beam. The change in the resonance frequency of a clampedclamped beam can be approximated using the Rayleigh’s energy method, 2

where ρ is the density and

(4) is the trial function [16]. Using 16

(5)

as the trial function and taking the derivative with respect to strain, the sensitivity can be approximated as [17].

(6)

Resonant MEMS strain sensors using DETFs as sensing elements have been investigated by various groups [9,10,12,16,17]. These devices can readily be fabricated using the semiconductor processes described in Section 3. The devices have demonstrated high resolution (less than 0.1 με in a 10 kHz bandwidth) which is beyond the performance of commercially available technology. Such devices have been characterized by bonding the underside of the substrate surface to a macro-scale mechanical structure such as a half-shaft [18]. This technology can be further developed to monitor drive-shafts of automobiles to monitor signatures from combustion chambers leading to automobiles with increased fuel efficiency. In addition, resonant MEMS strain sensors made from SiC have been studied to extend the operation to temperatures greater that 300○C enabling monitoring in extreme environments. This technology can also reduce the risk of brittle or corrosion failures of critical components used in oil and gas exploration or combustion systems. The temperature sensitivity of these devices should be considered and can be compensated with signal conditioning electronics.

5 Capacitive MEMS Strain Sensors Capacitive strain sensing is an alternative approach to resonant strain sensing, in which the strain is measured by monitoring the geometry [19] or permittivity [20]

MEMS Strain Sensors for In ntelligent Structural Systems

669

Fig. 5 Capacitiv ve strain sensor using electrostatic comb drives [26]

change in parallel plate capacitors. The passive nature of the capacitive straiin w power and has nominal temperature dependency [211]. measurement requires low Consequently, it requiress low voltage circuitry for continuous operation whicch makes it an attractive choice for industrial applications. MEMS capacitive straiin sensors generally consist of single or multiple capaccitors [22] or a system of electrostatic e comb drives [23]. The latter typically has a simple structure but requ uires amplification of the strain signals transferred from the medium by using sop phisticated circuitry or customized mechanical designs tto achieve acceptable rate off signal to noise ratio. In addition, a bulk micromachininng fabrication technique on semiconductor s grade wafers is typically used to manufaccture the capacitive senso ors and maximize the capacitive area. Therefore, thesse sensors are often fabricateed on top of relatively thick substrate. However, the preesence of a rigid substrate can dissipate up to half of the strain signal transferreed A a system of electrostatic comb drives (Figure 55) from the medium [24]. Although improves the capacitive area a [25] and gauge sensitivity, it will introduce readouut non-linearity, cross-axis strain s sensitivity, lower mechanical bandwidth, and largge parasitic and feed-through h capacitance into output signal. The capacitive sensing g method is based on the detection of electrical flux variations between two electriccally conductive electrodes when a differential voltage is applied across them. Thee amount of produced electrical flux (capacitance) deepends on the geometry of o the electrodes. For example, if we consider a paralllel plate capacitor (Fig. 6) th he capacitance, C depends on the area, A, the gap, g beetween the conductive plattes and permittivity of the medium, εoεr (Eq. 7).

C = ε oε r

A g

( A = bd)

(77)

Therefore, variation of thee physical dimensions will have an impact on the voltagge across the plates [27]. Furrthermore, parametric mechanical design is a critical steep in development of a capaccitive sensor.

70


Fig. 6 Schematic image of a parallel plate capacitor

Fig. 7 Unstrained capacitive strain sensor (top), translated and deformed state of capacitive electrodes due to applied strain (bottom)

One of the major challenges of capacitive strain sensor design is to eliminate the effect of the undesired strain signals and achieve accurate and reliable measurements. The majority of strain sensors are bulk micro-machined so the state of active strain field can be presumed as plane strain. As described in Fig. 7, the transferred plane strain will deform and translate the capacitive electrodes which influences on both capacitive gap (g) and area (A).Therefore, the capacitance change due to the applied strain can be derived as:

dC A ∂g 1 ∂A = −ε o ε r 2 ⋅ + ε oε r ⋅ dx ∂x ∂x g g

(8)

If we consider capacitive plates with the length, L, and unit heights, the deformation of the elastic plates can be modeled using the Euler-Bernoulli beam theory [28]:


EI

71

∂2y =M ∂x 2

(9)

where y, I and M are the beam deformation, moment of inertia, and applied moment, respectively. The initial boundary conditions for rotation (θ) and displacement for the unstrained (initial) and strained (deformed) beams are:

θ initial (x = 0, L ) = 0 , y initial (x = 0, L ) = 0

(10)

θ (1), deformed (x = 0 ) = θ 1 , y (1), deformed (x = 0 ) = Δ 1

(11)

Using (Eq. 9), the deflection of beams (1) and (2) are derived as:

y (1),deformed =

y( 2),deformed =

M2 2 EI

⎛ 2 x3 ⎞ M1 3 ⎟⎟ − ⎜⎜ x − x + θ1 x + Δ1 3 L ⎠ 6 EIL ⎝

M 3 ⎛ 2 x3 ⎞ M 4 3 ⎜ x − ⎟⎟ − x + θ1 x + Δ 3 + g o 2EI ⎜⎝ 3L ⎠ 6 EIL

(12)

(13)

Therefore, the capacitive gap change can be calculated by taking the derivative of the displacement with respect to its lateral displacement, y:

∂g = ( y( 2),deformed − y(1),deformed ) − g o ∂x

(14)

Moreover, the lateral displacement of the electrodes will change the capacitive area. Eq. 12 and Eq. 13 can be simplified by inclusion of symmetry planes and imposing specific boundary conditions on the mechanical design. On the other

300μm

Fig. 8 High resolution capacitive strain gauge designed to attenuate cross-axis strain signals [29]

72


hand, any strain measurement may be influenced by both strain components εx and εy (Eq. 8) which makes it extremely difficult to monitor the field when detection in a single direction is interested. Therefore, it is preferred to attenuate the cross-axis strain through mechanical design to eliminate the need for complex circuitry. In addition to structural symmetry, special arrangements of parallel plate capacitors such as differential readout can assist with the rejection of undesired mechanical signals. Therefore, capacitive sensors (Fig. 8) can be designed as a robust means to measure strain in small ranges (1 με to 1000 με) while exhibiting nominal signal attenuation due to cross-axis sensitivity [22].

6 Packaging and Data Acquisition As mentioned previously, the packaging of MEMS strain sensors is an important design consideration. This is due to the fact that the packaging itself could lead to device failure or performance drift. For example, if the sensor is not appropriately bonded or integrated into the material structure, delamination could occur leading to a failed sensor component or worse a failed structure. Furthermore, strain transfer from the mechanical structure to the sensor itself should be modelled and predicted to avoid signal biasing. In addition, to ensure long term operation, the sensor should be protected from the external environment with packaging. For instance, extreme temperature excursions, particles and humidity can cause sensor drift and sensor housing (in vacuum) can mitigate these issues. The electronics that interface with MEMS strain sensors can improve the performance of the sensor system. For example, the sensor itself converts mechanical deformations to electrical signals and those signals can be amplified or conditioned with external electronics. This leads to devices with signal drift compensation, thermal compensation and decreased noise. In addition, wireless telemetry of sensor data can be realized with the wireless sensor electronics which could lead to sensor networks and more streamlined systems. The packaging of the sensor and the electronics as an integrated system should be considered and developed.

7 Conclusion MEMS strain sensors (resonant and capacitive) have been developed and studied by various groups. These devices are fabricated using processes and materials developed by the semiconductor industry for IC technology. The processes enable batch fabrication leading to significant cost reductions. Ultimately, MEMS strain sensing can improve the performance of mechanical structures and systems with real-time monitoring. The efficiency, safety factors and operation lifetimes can be increased with sensors. In addition, wireless sensing architectures composed of MEMS strain sensors and sensor electronics can enhance the design of buildings, automobiles and aircrafts.


73

References 1. Bustillo, J., Howe, R., Muller, S.: Surface micromachining for microelectromechanical systems. Proc. of the IEEE 86(8), 1552–1574 (1998) 2. Akyildiz, F., Su, W., Sankarasubramaniam, Y., Cayirci, E.: Wireless sensor networks: a survey. Computer Networks 38(4), 393–422 (2002) 3. Senesky, D., Jamshidi, B., Cheng, K., Pisano, A.: Harsh Environment Sili-con Carbide Sensors for Health and Performance Monitoring of Aerospace Systems: A Review. IEEE Sensors Journal 9(11), 1472–1478 (2009) 4. Chang, F.: Structural Health Monitoring: Current Status and Perspectives. CRC Press, Boca Raton (1998) 5. Ghezzo, F., Rye, P., Huang, Y., Nemat-Nasser, S.: Integration of sensing networks into laminated composites. In: Proc. of SPIE Int. Symp. on Smart Structures and Materials (2008) 6. Hautamaki, C., Zurn, S., Mantell, S., Polla, D.: Experimental evaluation of MEMS strain sensors embedded in composites. Journal of Microelectrome-chanical Systems 8(3), 272–279 (1999) 7. Mall, S.: Integrity of graphite/epoxy laminate embedded with piezoelectric sensor/actuator under monotonic and fatigue loads. Smart Materials and Structures 11, 527–533 (2002) 8. PolyMUMPS Design Handbook, http://www.memscap.com 9. Azevedo, R., Jones, D., Jog, A., Jamshidi, B., Myers, D., Chen, L., Fu, X., Mehregany, M., Wijesundara, M., Pisano, A.: A SiC MEMS Resonant Strain Sensor for Harsh Environment Application. IEEE Sensors Journal 7(4), 568–576 (2007) 10. Myers, D., Cheng, K., Jamshidi, B., Azevedo, R., Senesky, D., Chen, L., Mehregany, M., Wijesundara, M., Pisano, A.: Silicon carbide resonant tuning fork for microsensing applications in high-temperature and high G-shock environ-ments. Journal of Micro/Nanolithography, MEMS, and MOEMS 8(021116) (2009) 11. Saponara, V., Horsley, D., Lestari, W.: Structural Health Monitoring of Glass/Epoxy Composite Plates Using PZT and PMN-PT Transducers. Journal of Engineering Materials and Technology 133(011011) (2011) 12. Wojciechowski, K., Boser, B., Pisano, A.: A MEMS Resonant Strain Sensor in Air. In: Proc. 17th IEEE International Conference on Micro Electro Mechanical Systems, pp. 841–845 (2004) 13. Azevedo, R., Zhang, J., Jones, D., Myers, D., Jog, A., Jamshidi, B., Wijesundara, M., Maboudian, R., Pisano, A.: Silicon Carbide Coated Silicon MEMS Strain Sensor for Harsh Environment Applications. In: Proc. 20th IEEE International Conference on Micro Electro Mechanical Systems, Japan, pp. 643–646 (2007) 14. Wojciechowski, K., Boser, B., Pisano, A.: A MEMS resonant strain sensor with 33 nano-strain resolution in a 10 kHz bandwidth. In: Proc. IEEE Sensors Conference, USA, pp. 947–950 (2005) 15. Young, W., Budynas, R.: Roark’s formulas for stress and strain, 7th edn., pp. 196–197. McGraw-Hill, New York (2002) 16. Roessig, T.: Integrated MEMS tuning fork oscillators for sensor applications. Ph.D. Thesis, Department of Mechanical Engineering, University of California, Berkeley (1998) 17. Wojciechowski, K.: Electronics for Resonant Sensors. Ph.D. Thesis, Department of Electrical Engineering, University of California, Berkeley (2005)

74


18. Myers, D., Pisano, A.: Torque Measurements of an Automotive Halfshaft Utilizing a MEMS Resonant Strain Gauge. In: Proc. of 15th International Conference on SolidState Sensors, Actuators, & Microsystems, USA, pp. 1726–1729 (2009) 19. Filanc-Bowen, T., Kim, G., Shkel, Y.: Novel Sensor Technology for Shear and Normal Strain Detection with Generalized Electrostriction. Proceedings of IEEE Sensors 2(4), 1648–1653 (2002) 20. Arshak, K., McDonagh, D., Duran, M.: Development of New Capacitive Strain Sensors Based on Thick Film Polymer and Cement Technologies. Sensors and Actuators A. 79, 102–114 (2000) 21. Cockbain, A., Horrop, P.: The Temperature Coefficient of Capacitance. British Journal of Applied Physics 2(9), 1109–1115 (1968) 22. Jamshidi, B., Azevedo, R., Jog, A., Pisano, A.: Silicon Cross-Axis Rejection Capacitive Strain Gauge. In: Proc. of ASME International Mechanical Engineering Congress and Exposition, USA (2007) 23. Guo, J., Kuo, H., Young, D., Ko, W.: Buckled Beam Linear Output Capacitive Strain Sensor. In: Proc. of Solid State Sensor, Actuator and Microsystems Workshop, USA (2004) 24. Azevedo, R., Chen, I., O’Reilly, O., Pisano, A.: Influence of Sensor Substrate Geometry on the Sensitivity of MEMS Micro-Extensometers. In: Proc. of International Mechanical Engineering Congress and Exposition, USA (2005) 25. Aebersold, J., Walsh, K., Crain, M., Martin, M., Voor, M., Lin, J., Jackson, D., Hunt, W., Naber, J.: Design and Development of a MEMS Capacitive Bending Strain Sensor. Journal of Micromechanics and Microengineering 16, 935–942 (2006) 26. Guo, J., Suster, M., Young, D., Ko, W.: High-Gain Mechanically Amplified Capacitive Strain Sensors. In: Proc. of IEEE Annual Meeting, pp. 464–467 (2005) 27. Jamshidi, B.: Poly-Crystalline Silicon Carbide Passivated Capacitive MEMS Strain Gauge for Harsh Environments. Ph.D. Thesis, Department of Mechanical Engineering, University of California, Berkeley (2008) 28. Azevedo, R.: Design and Evaluation of a MEMS Offset Capacitive Comb Strain Sensor. M.Sc. Dissertation, Department of Mechanical Engineering, University of California, Berkeley (2003) 29. Jamshidi, B., Azevedo, R., Wijesundara, M., Pisano, A.: Corrosion Enhanced Capacitive Strain Gauge at 370°C. In: Proc. of the 6th Annual IEEE Conference on Sensors, USA (2007)

A Pattern-Based Framework for Developing Wireless Monitoring Applications James Brusey1, Elena Gaura1, and Roger Hazelden2 1

Cogent Computing Applied Research Centre, Coventry University, Coventry, UK [email protected] 2 TRW Conekt, Solihull, UK [email protected]

Summary. Development of application-specific wireless monitoring systems can benefit from concept reuse and design patterns can form the enabling medium for such reuse. This chapter presents a set of five fundamental node-level patterns that resolve common problems when programming low-power embedded wireless sensing devices. Although the design patterns proposed are not subjected to a quantitative evaluation, a qualitative evaluation is performed through examining examples of these patterns in existing published deployments and systems. This analysis demonstrates that key deployment lessons are codified in each pattern.The pattern set forms a framework that is aimed at ensuring simple and robust deployed systems.

1 Introduction The need for distributed sensing for Structure Health Monitoring (SHM) is commonplace. Sensing both needs to be local to the phenomena (such as crack sensors on weld joints) and must cover a region (such as a structure with many weld joints). Distribution is often best served by wireless sensors: 1. They are quick to set up and tear down and the associated infrastructure is minimal. 2. Wireless sensors avoid the need for installing cabling for communication and power. 3. Individual sensing points can be added, moved, or removed at low cost. The idea of automatically and wirelessly acquiring data from a distributed set of sensors is relatively recent—feasible wireless sensors have only been readily available for the last decade or so. Although the technology is beginning to move into the mainstream, developers are still faced with a technology that is hard to understand and difficult to make reliable. A key difficulty, for example, is in powering sensors and wireless transmitters, thus optimising energy efficiency of a Wireless Sensor Network (WSN) can often be critical to the business case for their use. For SHM in aircraft, a minimum battery lifetime of about seven years is required to ensure that the wireless system is cost effective. S.C. Mukhopadhyay (Ed.): New Developments in Sensing Technology for SHM, LNEE 96, pp. 75–91. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

76

J. Brusey, E. Gaura, and R. Hazelden

Where distributed sensing is used for one application, there often arise multiple applications for the sensory data and thus the need for sensing and sensor-based actuation evolves over time. For this reason, a well-structured, systematic development and framework are required to ensure that new applications and additional sensors can be easily integrated as the system grows. Some support exists for simplifying the development of wireless sensors (e.g. TinyOS, Embedded Linux). However, there is little work on guidelines or frameworks that establish best practice in this area. In particular, there appears to be a series of identifiable lessons that are being repeatedly rediscovered by programmers and research groups. Part of the difficulty is that WSN deployments are diverse. There appears to be little carry-over in terms of lessons learnt from one deployment to the next simply because many of the issues do not apply. This chapter sets out a framework for reusable patterns for WSNs. Although no quantitative evaluation of the framework is provided here (and indeed may not be possible), a qualitative evaluation is performed by identifying key lessons from several deployed systems and linking these to elements of the framework. This work builds on prior work [4] that formulated the framework mathematically. Although precise, the mathematical formulation may be less intuitive for some readers. In comparison, it is hoped that this presentation is more accessible while the mathematical formulation can be used for reference where clarification is needed. The presentation here is loosely based on design patterns [6] but without the emphasis on object-orientation. The next section presents the framework and the associated five node-level patterns: the Filter pattern, the Event Detector pattern, the Priority Buffer pattern, the Nonpreemptive Scheduler pattern, and the Interval Listening pattern. Each pattern is described in terms of its aims, the triggers that indicate the need for it to be considered, collaborations that can occur with other patterns, possible extensions, and finally examples of pattern usage from the literature. The final section concludes the work and outlines how it might best be applied to aiding the development process of novel monitoring applications.

2 A Pattern-Based Framework In this work, the term “framework” is used partly to refer to a form of protoarchitecture where elements may be added, removed, or altered to suit an application and partly to encapsulate the collection of related design patterns. The design pattern literature is relatively well-established as a means of describing prototypical solutions for common object-oriented design problems. This work borrows the term “design pattern” but without the object-oriented undertones and associated requirements. Here, a design pattern is intended to be merely a template or guide for solving specific design problems with software development for WSN systems. The framework presented here is based on two fundamental assumptions: 1. There are benefits to processing at the node in a wide variety of applications. 2. Benefits from sharing information between (leaf) nodes within a network are rare in practice and even less frequently worth the associated risks and complexity.

A Pattern-Based Framework for Developing Wireless Monitoring Applications

77

sense

filter

interval listening non-preemptive scheduler

event detector priority buffer

transmit

Fig. 1 Pattern-based Framework for nodes. A dashed box is drawn for elements that have no explicit pattern described here

This framework, then, focuses on a simple WSN where remotely stationed wireless nodes communicate directly (single hop) or indirectly (multi-hop) with a “sink” or base station. This approach is aligned with Raman and Chebrolu [15] who argue that the WSN domain is divided into two main camps: Those devoted to devising algorithms and protocols and those pursuing application centric design and deploying systems “in the wild”. They find that the sophisticated algorithms and protocols devised by the former group are rarely used by the latter group. Deployed systems tend to eschew complexity. Simpler, well understood MAC layers are favoured. Werner Allen et al. [17] in their description of Lance also argue for keeping node interaction simple. In their volcano monitoring system, sufficient complexity (and frustration!) was introduced even by using the well known FTSP time synchronisation protocol, which at the time, had some unresolved bugs. The Lance architecture specifically assumes that remote nodes do not collaborate or share information. Comprehensive arguments such as those above lead to the key elements of the node-level framework, which are summarised in Figure 1. The central task for the node begins with the “sense” operation. Noise is filtered and the original data is transformed into meaningful information. Event detection occurs next potentially leading to a message that must be transmitted. The message is buffered according to priority before being transmitted. These tasks must be interleaved along with sleeping and listening at intervals and this is the job of the scheduler. It is usual, when developing a monitoring system, to include various data structures (or perhaps classes) that are specific to what is being measured. This framework describes the sensed data and the subsequent inferred state as simple vectors. The framework does not preclude using more complex structures but flattening the data structure in this way emphasises the generality of the approach to a wide variety of possible sensors, phenomena, and applications.

78


In the following sections each design pattern is described in more detail. A general description of each pattern is given, followed by its Aims (what is the intended function achieved by the pattern), Triggers (the reasons why the pattern might be used), Collaborators (which other patterns are often used with it), Possible Extensions (additional functionality) and Examples (some examples from the literature of successful use of the pattern). 2.1 Filter (or Model-Based Smoothing) Pattern The Filter Pattern, which could also be termed the Model-based Smoothing Pattern, is a node-level pattern to smooth the raw, sensed data and / or infer the state of the phenomena at the node. There are some specific reasons for performing filtering at the node rather than at the sink, as described in Section 2.1.2. Thus this approach is advantageous in some applications and not others. The Filter Pattern is based roughly on the design of a Kalman Filter (see Welch and Bishop [16] for a introduction) and its structure is summarised in Figure 2. This is a recursive, on-line filter that takes as input a vector representing the estimate of prior state, the elapsed time, and a vector containing current sensor readings and produces as output an updated state estimate. Kalman Filters may be too computationally intensive or difficult to implement (due to the requirement for floating point arithmetic and matrix inversion) for most WSN applications, however the overall design is still applicable. A light weight alternative, which has a similar recursive structure, is the Exponentially Weighted Moving Average (EWMA) filter.

sensor readings Filter

new state vector

last time last state vector Fig. 2 Synopsis of the Filter pattern

2.1.1 Aims A filter aims to do the following, based on a series of sensor readings: • Reduce noise, and / or, • Summarise a sensory “chunk”, and / or, • Infer state, possibly from sensors of differing modality.

Most sensor measurements incorporate some form of noise. The effect of noise can often be reduced by applying some form of low-pass filter such as an EWMA or Kalman Filter. Although the computation involved is typically more easily performed at the sink, filtering at the node helps to ensure that other processing, such as event detection, is affected minimally by noise.


79

Filters can also summarise a “chunk” of sensor measurements. A common application is processing audio data, where a large amount of data is being received by the sensor during each sampling period. Summarisation can substantially reduce the bandwidth required as the summaries are generally much smaller than the original data. Summarisation might be generic (such as finding the peak audio levels over the last second) or specific (is the call of a particular animal species being heard?). Summarisation might be final (a cane toad call is being heard) or tentative (possible earth tremor occurring currently). In the latter case, the unfiltered data might need to be stored for later analysis, perhaps when other nodes have confirmed that a tremor event occurred. Note that it is important to keep distinct the two issues of: 1. filtering, which transforms data into an estimate or summary of the state, and 2. event detection, which detects whether the change in the state is significant. Such separation ensures that the two distinct concerns of deriving a state representation and decision making based on that state are not confused. Filters are often explicitly model-based, in which case, the filter attempts to derive an estimate of the state of the system based on past sensor readings. Often it is possible to assume that the system has the Markov property, which means that the most recent sensor reading and the last state estimate are all that are required to estimate the current state and that no better can be done by knowing the complete history of states. For example, when estimating the number of people in the room from a door sensor, it is only necessary to know the past state (how many people were previously in the room) and how many people exited or entered. A model-based approach is often useful to fuse data from several sensors of differing modalities. For example, there may be a Passive Infra-Red (PIR) sensor detecting overall occupancy alongside the exit sensor attached to the same node1 . Information can be combined from the pair of sensors to give a more accurate estimate of the true occupancy. 2.1.2 Triggers The Filter Pattern should be used when: • Available bandwidth is low relative to the amount of data sensed. • The relative cost of transmitting data is higher than processing it on the node • Actual sensor readings are not necessarily required (or only contingently

required). Many wireless sensor deployments involve a set of individual sensors that in combination can provide much more data per unit time than the wireless transmission medium can sustain. High data rate sensing applications involving video, audio, or vibration sensors, in particular, are limited by the available wireless bandwidth. Such situations provide a strong motivation to attempt to process and summarise 1

Fusion between sensors on different nodes might also be possible but the framework encourages us to first consider moving such sensor fusion to the sink.

80


sensor readings on the node. Further reductions in bandwidth requirements are possible by combining the Filter pattern with the Event Detection pattern. Transmission costs tend to be an order of magnitude greater than the computational costs associated with filtering. Listening costs can also be a significant factor, however, if the node needs to act as a router (which is often needed in mesh networking). Individual sensor readings may not be required. For example, an application may involve monitoring cupboard “opening” and “closing” events using a light sensor. It is not necessary to know how much light is falling on the sensor when the door is open, it is enough to know that the door is open. Filtering the data into a state variable of “open” or “closed” not only reduces the needed bandwidth but also better supports subsequent Event Detection by reducing the effect of noise and thus reducing spurious event detection. 2.1.3 Collaborators The Filter pattern is often used in collaboration with the Event Detection pattern. There are several reasons why such a collaboration can be useful: 1. The process of filtering removes noise thus reducing spurious event detection. 2. Transforming the raw sensor data into a state vector simplifies the task of identifying whether the state has changed in a way that can be considered a meaningful event. 3. It supports avoiding a “slippery slope” problem where the event detection mechanism cannot detect a change if the change occurs slowly enough. Further, combining the Filter pattern with the Interval Listening pattern can avoid the possibility that the energy gained from reduced transmissions is not then subsequently lost due to increased listening time. 2.1.4 Possible Extensions Although the framework begins with the assumption that individual nodes do not share information and are not required to communicate directly with one another, for some applications, it may be useful to allow such communication. In this case, the state estimate produced by the Filter can take into account measurements from neighbouring nodes. For example, an intrusion detection system might consider a possible intrusion more likely if neighbouring nodes are also sensing a disturbance. The Lance architecture suggests a useful extension that involves locally storing the original (unfiltered) data and providing it on-demand. This approach is particularly useful when local information is not sufficient to fully make a decision about how useful the data is. The state vector need not be just about the phenomena. It is often useful to include management information or, in other words, information about the state of the sensors or the wireless node. For example, this could include local timestamps, battery voltages, estimates of uncertainty in measurement readings, link reliability statistics, and so forth.


81

2.1.5 Examples On-node filters are quite common in the literature. Here are two examples: • Lance [17] is an architecture built originally for volcano monitoring that made

use of audio sensors. Transmitting all of the audio over a multi-hop network led to much contention and low yield. By sending summaries of the data instead, the bandwidth requirement was significantly reduced and the yield of useful data improved. • The Cane Toad monitoring project [10] is another excellent example of successfully processing complex audio data on the node. Frog calls were collected in the wild and analysed in real-time using spectrograms and C4.5 decision trees to classify the frog species. The initial work required more sophisticated processors but the final systems were able to be deployed on Mica2 motes. Performing the conversion from raw audio to identified frog species on-board was critical in reducing the bandwidth requirement for this application and thus allowing it to be performed with inexpensive motes. 2.2 Event Detector Pattern A fundamental and simple pattern for reducing the number and size of transmissions is based on detecting events at the node. The pattern is summarised in Figure 3. Event detectors typically work by comparing the current state with the last transmitted state. If the difference exceeds some threshold, then an event is detected. Comparing with the last transmitted state avoids the possibility of sending duplicate event messages while using the last transmitted state for comparison avoids a “slippery slope” effect where a slowly changing phenomena may appear to be uneventful (the gradient at any point is low) but the long term change is still significant.

current state

Event Detector

event detected?

last transmit time last transmitted state Fig. 3 Synopsis of the Event Detector Pattern

2.2.1 Aims The Event Detector Pattern aims to: • reduce the transmission of unnecessary data • allow for increased rate of transmission of needed data

When performing initial investigation, it is often useful to acquire a large number of samples from a large array of sampling points. Once this initial phase has ended

82


and the phenomena is better understood, much of this data is not needed and continuing to acquire it is an unnecessary cost. Identifying which data is unnecessary will depend on the application however there are some common cases: 1. where the state maintains the same (or nearly the same) value over an extended period; 2. where the state is easily predictable over time (e.g. a linear trend); or, 3. where the state continues to stay within a set of expected values (but it is not necessary to know exactly which). The term “state” is used above, rather than “sensor reading” since it is typically the case that the raw, unsmoothed, uncalibrated reading will first be processed into an application-specific state vector by a Filter prior to event detection. For example, it is simpler to design event detection based on a state vector that includes, say, an estimate of the residual life rather than one that gives wear sensor measurement readings. To properly detect events for systems that are predictable over time, some prediction is needed. For example, if the last transmitted state was taken 5 minutes ago and indicated that the state was at 1 unit and rising linearly by half a unit every minute, then the predicted state is 3.5 units. If the new state estimate is close to this (within some threshold), then it is considered uneventful. In principle, arbitrarily complex models could be used here. In practice, however, simple linear regression is sufficient for most cases. A further advantage of event detection is that it may save sufficient transmission energy and bandwidth to allow an increase in sensing frequency. This potentially allows detection of short-lived phenomena that might be missed otherwise. 2.2.2 Triggers The Event Detector pattern should be considered when: • the system being measured has a steady or easily predicted state for extended

periods, and / or, • transmission cost (say, in terms of energy or bandwidth use) is high.

Steady state systems are reasonably common and, for these systems, the use and benefit of the Event Detector pattern is more obvious. Less obvious is the application of event detection to systems that follow diurnal, periodic, or short term linear trends. Some examples include: temperatures within a building, water pressure within the water supply pipe network, wear on machine bearings, and so forth. 2.2.3 Collaborators The Event Detector Pattern is often used in conjunction with the Filter Pattern (also see section 2.1.3). In fact, they are so often used jointly that it is easy to confuse them or not to know when to use one without the other. Filters are used without event detection when the decision about whether or not an event has occurred must be deferred until more information is known. Perhaps


83

the decision can only be made at the sink, when summaries from other nodes have been heard. Event detection is used without filtering when the sensor already provides a sufficiently clean signal. For example, an RFID tag reader provides tag-read messages that are free from noise. Event detection is needed to identify when tagged items appear or disappear. Even in this case, it may still be useful to have a “filter” to organise the incoming tag-read messages into an estimate of which items are present (i.e. a representation of state). 2.2.4 Possible Extensions There are several ways in which to extend the basic Event Detector pattern: • Incorporating a “heartbeat” message can ensure that the sink will eventually de-

tect node failure. Without this, the node might not send any data for an indefinite period, if the phenomena is in a steady state. A simple method to incorporate a heartbeat is to signal an event if the last transmission time was long ago, even if the state is unchanged. (The exact definition of how long to wait before sending a heartbeat will depend on the application.) • Model-based event detection (based on predicting from linear or other trends) can be further enhanced by assuming that the sink can also apply the same model-based prediction. The Spanish Inquisition Protocol [8] describes an event detector that makes use of dual prediction (on both node and sink). • A useful assumption is that the state vector (used as input) has the Markov property. That is, a prediction based on the state vector would not be improved by knowing the complete history of states. This is a helpful consideration when deciding what features to include in the state vector. For example, rate of change is needed if one wants to predict based on a linear extrapolation of the trend. 2.2.5 Examples The use of the Event Detector pattern is commonplace in the literature. Two interesting examples both begin with high data rate sensors. • VoxNet is a deployed WSN that localises animal calls using a set of four mi-

crophones at each node [2,1]. Full trilateration of incoming audio signals could only be performed at the sink, however sending all of the audio signals tended to overload the 802.11 network used. Setting up an event detector that looked for the start and end times of possible animal calls meant that audio data could be sent much less frequently. • Event detection for human activity monitoring systems can substantially reduce transmissions. In work elsewhere [3], a postural activity monitoring system was developed that classified posture based on 2 or more MEMS accelerometer sensors worn on various places on the body. A combination of on-board posture classification, an exponentially weighted voting filter and event detection reduced the transmissions from the original 10 Hz to about 600 event transmissions in 30 minutes (0.3 Hz) without the filter or around 100 in a 30 minute period with filtering included (0.06 Hz).

84


2.3 Priority Buffer Pattern A synopsis of the Priority Buffer pattern is given in Figure 4. The priority of any message is determined by the contents of the message. This priority is used to determine ordering in the priority buffer along with the time of arrival. Transmission then proceeds from the head of the buffer.

transmission

current state xt

calculate priority p (xt )

p (xt )

xt

.. .

.. .

Fig. 4 Synopsis of the Priority Buffer Pattern

This simple pattern can be critical in ensuring that high priority messages are communicated successfully. The first part of the pattern consists of ordering the buffer according to priority. The second part consists of controlling the timing of transmission and, in particular, controlling when transmissions should be retried. 2.3.1 Aims The aim of the Priority Buffer pattern is to: • increase the likelihood that important packets are transmitted, • reduce the contention for the transmission medium, • gracefully handle extended periods without the ability to transmit.

Traditional wired networks assume that the probability of any given transmission failing is always the same. Wireless networks, however, suffer from variations in failure probability. For example, mobile wireless devices may be in range and able to communicate for some long period and then subsequently out of range or RF occluded from communicating for a period. Fixed devices can have similar variations in failure probability due to environmental factors such as rain or snow, the movement of occluding objects, and so forth. For this reason, when a transmission fails, particularly if it has already failed several times, it may not be best to retry immediately.


85

The Priority Buffer pattern responds to this problem by: 1. raising the communication buffer to an application level (rather than an operating system one), 2. allowing re-ordering of transmissions by priority even when a transmission has failed, and, 3. allowing control of when retries should occur. Dealing with the communication buffer at an application level means that it is possible to support much larger buffers than usual, perhaps making use of flash memory. Furthermore, since message headers have not been added yet, the individual messages will be smaller. When communication is cut for an extended period, this application buffer may be sufficient to ensure that no information is lost. Some communication messages are of high priority (“the building is on fire”) while others are of low priority (“the temperature in the hallway was 20 °C”). Priority may also depend on time (“temperature was x, three hours ago” may be less important than “temperature is currently y”). If communication has been prevented for several hours, when communications returns it makes sense to ensure that high priority messages are sent first. Sending them first will also mean that it is more likely that they are transmitted successfully. Where communication is failing because of contention for the transmission medium, reducing the number of attempts to transmit will help to reduce contention and this is an important consideration for the designer of a Priority Buffer. Communication may also be failing due to a transient environmental effect (such as rain or snow) that will continue to prevent successful transmission for some time. An application-level strategy can balance the importance of timely transmission against the cost of many retries. 2.3.2 Triggers The Priority Buffer pattern should be considered when: • some messages are of higher priority than others, • the likelihood of successful transmissions varies over time.

2.3.3 Collaborators A critical question when devising a Priority Buffer is how to determine which messages are more important. In particular, the message should contain enough information to enable a decision about its priority to be made. This implies an interaction with the associated Filter. The Filter helps the Priority Buffer by placing sufficient context into the state vector. For example, in a patient health monitoring application, rather than just stating the vital signs (heart rate is 180, core temperature is 39 °C, etc), the Filter can help by interpreting the data somewhat (“likelihood of imminent cardiac arrest is 30%”). In principle, the burden of inference can lie in either place. In practice, however, it can be easier to encode the priority decision as a set of rules. Also, this ensures that any computed inference is encoded as part of the state vector and thus transmitted to the sink.

86


2.3.4 Examples During the development of a glacier monitoring application, Martinez et al., [11] had the problem of wirelessly transmitting from the glacier to an Internet café several miles away. To save power, communication was reduced to a few transmissions per day. However, snow storms would severely disrupt communication. If the transmission was continuously retried throughout the storm, it would just drain the batteries. Therefore, a series of 3 failures caused the node to give up transmitting for several hours before retrying. The above example illustrates how it is important to consider the application and its environment. It also shows how useful it is to elevate the question of when to retry to an application-level, rather than leaving this to the operating system, to avoid wasting battery power and to allow consideration of the priority of the message being communicated. 2.4 Nonpreemptive Scheduling Pattern Nonpreemptive scheduling is a central component of simple, embedded operating systems such as TinyOS. There are two reasons for declaring this as a pattern. The first reason is that an understanding of the implications of this pattern will help developers understand how to best use TinyOS and similar systems. The second reason is that there are still many specialised applications that call for simpler hardware or more stripped down software than TinyOS or a similar system would allow but where task interleaving and timed operations are still required. A synopsis of the Nonpreemptive Scheduling Pattern is given in Fig. 5. The scheduler maintains a “step schedule” or list of active “steps” and their associated start times. A step is a short-lived task. For example, beginning to send a message is a step, whereas the whole process of sending a message is a series of steps. The

Step schedule

at t1

before t1

sleep until t1

run step s1

s1

t1

s2

t2

.. .

.. .

update schedule and reorder

Fig. 5 Synopsis of the Nonpreemptive Scheduler Pattern (adapted from Cassandras and Lafortune [5])


87

schedule process begins by taking the first item from the list of steps and sleeping until its start time. If the start time has arrived, the step is executed. If the start time has not yet arrived, it may be due to waking for another reason such as an interrupt (and thus the schedule may need updating). The last step is to update the schedule and reorder according to start time. Updating the schedule involves asking each possible task whether they can (still) perform a step. While preemptive scheduling is the norm in modern computers, low-power processors such as Texas Instrument’s MSP430 or generic PIC processors, which are widely used for WSN applications, are limited in this regard. Nonetheless, hardware interrupts, due to timers and I/O, will interrupt other tasks and care is needed to ensure that there are no race conditions as a result for any buffers shared between the main process and interrupt routines. In most programming idioms, each subroutine or module, once started, will run to completion. A program that calculates π to one million decimal places will hold the CPU captive for as long as the task requires. Multi-tasking operating systems avoid this problem by preemption. That is, they interrupt the task, save its state, and switch to a new task transparently. This allows other tasks to carry on working while the calculation is ongoing. preemptive multitasking, however, is expensive (in terms of memory and CPU overhead) and may be difficult to support on low-power processors. 2.4.1 Aims The Nonpreemptive Scheduling pattern aims to: • provide efficient interleaving of sleeping, sensing, listening, and transmitting

cycles, • allow for timed communication for listening and transmitting, • support long running or slow external sensors with minimal CPU

One way to allow fine-grain interleaving of tasks without preemption, is to recode each module as a state machine. A simple example is shown below in terms of the original code in Figure 6 and the equivalent state machine in Figure 7. Note that recursive routines first need to be converted to iterative ones prior to conversion. simple-stmt; if (cond1) stmt2; else stmt3; while (cond2) stmt4; Fig. 6 Original procedural psuedo-code

88

J. Brusey, E. Gaura, and R. Hazelden int state=1; bool fsm1_feasible() { return state != 7; } time_t fsm1_when() { return now(); } void fsm1_step() { if (state==1) { simple-stmt; state++; } else if (state==2) { if (cond1) state=3; else state=4; } else if (state==3) {stmt2; state=5;} else if (state==4) {stmt3; state=5;} else if (state==5) {if (cond2) state=6; else state=7; } else if (state==6) {stmt4; state=5;} else state = 7; } Fig. 7 Reworked code as a state machine

For the example in Fig. 7, the finite state machine method fsm1_step() must then be called repeatedly (by the scheduler) until it is no longer feasible (specifically, fsm1_feasible() returns FALSE). Each method call instance is termed a step. This structure allows an arbitrarily large set of tasks to be interleaved without preemption. Recoding as a state machine solves the problem of interleaving but to support operations that are triggered on a timer or a hardware interrupt, the scheduler must keep track of the list of pending tasks along with their associated execution time, ordered by earliest time first. Sleeping can occur until the earliest task time is reached or a hardware interrupt occurs. Cassandras and Lafortune provide a detailed example of the workings of such a scheduler in the context of discrete event simulation [5]. Correctly dealing with hardware interrupts is a key issue for this pattern. As pointed out by Pont [14], high priority interrupt service routines may mask lower priority interrupts from being serviced. Therefore, interrupt service routines must be minimalist—perhaps even just waking and setting a flag to note their occurrence. The scheduler must therefore be prepared to be woken before the next schedule step time and in this case, update the schedule according to which steps are now feasible. For example, imagine a button press event handler OnButtonRelease. The button press triggers a hardware interrupt, possibly waking the system from sleep, that records that a button press has occurred. At this stage, the OnButtonRelease_feasible method will return “true”. The schedule will be subsequently updated and the step method called until no longer feasible. 2.4.2 Triggers • Multi-tasking operating system avoided or not feasible • Need for more complex task interleaving than possible with a simple sense-

process-send-sleep cycle • Not possible to use an off-the-shelf nonpreemptive OS, such as TinyOS or Con-

tiki, perhaps due to limits of microprocessor, or in an attempt to reduce the power budget.


89

2.4.3 Examples The best known example of this type of scheduling pattern in the WSN domain is TinyOS. There are a number of other systems that use a similar approach (such as the JACK agent programming environment [9]). A common approach is to automatically rewrite the programmed code as a state machine (this is true for both NesC in TinyOS and the JACK compiler). The Nonpreemptive Scheduler pattern derives much of its design from Pont’s patterns for time-triggered architectures [14] and Cassandras and Lafortune’s description of timed automata [5]. 2.5 Interval Listening Pattern 2.5.1 Aims The Interval Listening pattern has the following aims: • support mesh-networking (most or all nodes act as routers), • allow nodes to spend most of their time asleep but still not miss (most) mes-

sages, • reduce the amount of time spent “idle listening”.

To function as a mesh network, individual nodes must be capable of acting as routers. In principle, this means that they must be ready to receive messages at any time. In practice, such high alertness is generally only required when nodes are initially deployed or subsequently moved. For most installations, communication quickly stabilises into a predictable pattern based on regular sensing cycles and well established routing paths. Therefore, despite the need for nodes to act as routers, they can predict when the next message will arrive and revert to an ultra-low power mode until then. 2.5.2 Triggers • Not all nodes are within communication range of the base station (and so mesh-

networking is required) • Nodes do not need to be actively sensing all the time

2.5.3 Examples One form of the Interval Listening pattern is implemented as Low Power Listening (LPL) [12]. This protocol is a simple extension of standard TinyOS message transmission. It works by attaching a one second long prefix to transmitted messages. The receiver then only needs to wake up once per second to check for any transmissions. This simple modification substantially extends the life of each node. The Time Synchronised Mesh Protocol (TSMP) [13] developed by Dust Networks is another approach to Interval Listening that is based on a combination of Time Division Multiplexing (TDM), where each node has a specific slot when it

90


can transmit, and integrated time synchronisation that works by replying back to any sender how late or early their packet was. Note that the integrated time synchronisation is needed for two reasons: (a) to ensure that nodes wake at the right time to listen to neighbours, and (b) to avoid the need for top-down time synchronisation. TSMP is potentially much more efficient than LPL since transmissions can be short and the node does not necessarily need to wake up as frequently as every second. TSMP is implemented in WirelessHART and is part of the ISA100 standard.

3 Conclusions The WSN domain is transitioning in much the same way as Computer Science transitioned towards Software Engineering in the past: from being a research-only domain that focused on optimising algorithms to being one that included a greater focus on the problem of developing reliable, functionally correct, useful and applicable systems. This naturally includes a greater consideration of the task facing WSN developers. Design patterns have revolutionised the way software is engineered; a similar revolution is needed in WSN engineering. The framework and patterns presented here should be taken as a work in progress. The WSN field continues to evolve. Nevertheless, they should also not be seen as untested. Examples throughout have shown that these patterns appear repeatedly in reports on functioning deployed systems. The framework presented here is only one in a range of possible design frameworks. Some applications will require extremely simple nodes that cannot perform any processing on-board, while others will be able to incorporate much more sophisticated algorithms and patterns than those suggested here. Nonetheless, it is our strong belief that this framework will provide useful guidance to WSN developers across a broad range of applications.

References 1. Allen, M., Girod, L., Newton, R., Madden, S., Blumstein, D.T., Estrin, D.: Voxnet: An interactive, rapidly-deployable acoustic monitoring platform. In: International Conference on Information Processing in Sensor Networks, IPSN 2008, pp. 371–382 (2008) 2. Allen, M.: VoxNet: Reducing latency in high data-rate applications. In: Gaura, et al [7] 3. Brusey, J., Gaura, E., Rednic, R.: Classifying transition behaviour in postural activity monitoring. Sensors & Transducers Journal 7, 213–223 (2009), http://www.sensorsportal.com/HTML/DIGEST/P_SI_98.htm 4. Brusey, J., Gaura, E.I., Goldsmith, D., Shuttleworth, J.: FieldMAP: A spatio-temporal field monitoring application prototyping framework. IEEE Sensors 9(11) (November 2009), http://dx.doi.org/10.1109/JSEN.2009.2021799 5. Cassandras, C.G., Lafortune, S.: Introduction to Discrete Event Systems. Kluwer Academic Publishers, Dordrecht (1999) 6. Gamma, E., Helm, R., Johnson, R., Vlissides, J.: Design patterns: elements of reusable object-oriented software. Addison-Wesley Professional, Reading (1995) 7. Gaura, E.I., Girod, L., Brusey, J., Allen, M., Challen, G.W. (eds.): Wireless Sensor Networks: Deployments And Design Frameworks (Designing and Deploying Embedded Sensing Systems. Springer, Heidelberg (2010)


91

8. Goldsmith, D., Brusey, J.: The Spanish Inquisition Protocol: Model-based transmission reduction for wireless sensor networks. In: Proc. IEEE Sensors. IEEE, Los Alamitos (2010) 9. Howden, N., Rönnquist, R., Hodgson, A., Lucas, A.: JACK intelligent agents - summary of an agent infrastructure. In: Proceedings of the 5th International Conference on Autonomous Agents (Agents 2001) (2001) 10. Hu, W., Bulusu, N., Dang, T., Taylor, A., Chou, C.T., Jha, S., Tran, V.N.: Cane toad monitoring: Data reduction in a high rate application. In: Gaura, et al [7] 11. Martinez, K., Hart, J.K.: Glacier monitoring: Deploying custom hardware in harsh environments. In: Gaura, et al [7] 12. Moss, D., Hui, J., Klues, K.: Low power listening. Technical Report TEP 105, TinyOS Core Working Group (2007) 13. Pister, K.S.J., Doherty, L.: TSMP: Time synchronized mesh protocol. In: Proc. IASTED Intl. Symp. Distributed Sensor Networks (DSN 2008), pp. 391–398 (2008) 14. Pont, M.J.: Patterns for time-triggered embedded systems: building reliable applications with the 8051 family of microcontrollers. ACM Press/Addison-Wesley Publishing Co. (2001) 15. Raman, B., Chebrolu, K.: Censor networks: A critique of “sensor networks” from a systems perspective. ACM SIGCOMM Computer Communication Review 38(3), 75–78 (2008) 16. Welch, G., Bishop, G.: An introduction to the Kalman filter. Technical report, University of North Carolina at Chapel Hill (1995) 17. Allen, G.W., Dawson-Haggerty, S., Welsh, M.: Lance: Optimizing high-resolution signal collection in wireless sensor networks. In: Proc. 6th ACM conference on Embedded Network Sensor Systems (SenSys 2008), pp. 169–182. ACM, New York (2008)

Distributed Brillouin Sensor Application to Structural Failure Detection F. Ravet Omnisens, Morges, Switzerland

Abstract. Disaster prevention in civil infrastructures requires the use of techniques that allow temperature and strain measurements in real time over lengths of a few meters to tens of kilometres. The distributed Brillouin sensor technique has the advantage to combine all these characteristics. The sensing mechanism of the DBS involves the interaction of two counterpropagating lightwaves, the Stokes and the pump, in an optical fibre. Spatial information is obtained through time domain analysis. An analytical model describing the sensing mechanism based on stimulated Brillouin scattering (SBS) interaction is introduced and validated experimentally. This model development leads to the implementation of a signal processing method grounded in the physics of Brillouin scattering. An analytical approximation, valid for the optimum sensing region, reconstructs the Brillouin spectrum distribution from input sensing parameters and measured data. These data are obtained with a spectrum analysis methodology, based on three original tools: the Rayleigh equivalent criterion, the length-stress diagram, and the spectrum form factors. This methodology has been successfully used on experimental spectra. The DBS and the signal processing approach were then used to monitor the structural changes of steel pipes, composite columns and concrete elements. The DBS measured the strain distribution of those structures while they were stressed. The DBS provided detailed information on the structure’s health at local and global level, associated with deformations, cracks and buckling. This work demonstrates that the DBS is capable of extracting critical information useful to engineers: engineer’s experience and judgement in conjunction with appropriate data processing methods make possible to anticipate structural failures.

1 Introduction Structural health monitoring (SHM) is an important development of the civil engineering field. Recent events remind us that the cost of a bridge collapse is much more than economical; it touches people’s life, as dramatically demonstrated by the collapses of the Laval’s La Concorde overpass, in fall 2006, and, the Minneapolis bridge, in summer 2007. The implementation of comprehensive SHM in civil infrastructure is made possible by the use of optical fiber sensors and can substantially improve the safety of civil structures and help to manage them more efficiently. More specifically, distributed Brillouin-based sensing systems (DBS) are capable of measuring strain everywhere along a dedicated optical fiber cable S.C. Mukhopadhyay (Ed.): New Developments in Sensing Technology for SHM, LNEE 96, pp. 93–136. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

94

F. Ravet

attached to the structure to be monitored. Measurement readings can be taken every tenth of cm, which is a clear advantage when localized faults, such as cracks, need to be detected, but their position is not known beforehand. Among the principal aims of SHM, the detection and localization of localized failures such buckling and cracks appear as an essential task as they can lead to lethal faults for the structure’s life. The present work introduces a technique combining high resolution distributed strain measurements based on Brillouin scattering and advanced signal processing methodology to detect localized structural failures. The approach is validated by laboratory trials and field experiments.

2 Distributed Brillouin Sensor General Description a. Brillouin frequency relationship with strain and temperature

The Brillouin effect is the scattering of a lightwave, called pump, by an acoustic wave [1]. In other words, the optical wave is scattered by a propagating periodic variation of the density of the medium. The scattered beam optical frequency experiences a Doppler shift known as Brillouin frequency shift (νB) which is expressed as νB =

2nV A

λp

.

(1)

According to Eq.(1), the Brillouin frequency is proportional to the refractive index of the fiber (n) and to the acoustics wave velocity (VA). It also depends on the pump signal wavelength (λp). Typically, the Brillouin frequency of ITU G.652 fibers [2] is about 12.80 GHz at 1.31 mm and 10.85 GHz at 1.55 μm. Figure 1 presents a typical Brillouin spectrum which is the product of a scattered lightwave at 1554nm.

Normalised Intensity

1.00 0.75 0.50

FWHM or

B

0.25 0.00 10750

B

10800

10850

10900

10950

11000

Frequency [MHz] Fig. 1 Thick curve: Backscattered Stokes spectrum from an incident lightwave (λp = 1554 nm) in a 20 km long ITU G.652 fiber obtained by a heterodyne method (Nazarathy 1989, Derickson 1998); the incident power is 310 μW; Measured Brillouin frequency shift is 10873.6 MHz and linewidth is 31 MHz. Thin curve: Lorentzian distribution drawn with the measured Brillouin frequency and linewidth.

Distributed Brillouin Sensor Application to Structural Failure Detection

95

A sound wave propagates in bulk silica [3] and in optical fibers [4] over a few microns only, their attenuation coefficient (~ 106 m-1) being much larger than the optical attenuation coefficient in the telecommunication window (~ 10-4 m-1). The acoustic wave decay can not be neglected as it determines the Brillouin spectrum linewidth. It is commonly accepted in the literature that the decay follows an exponential law [5]. The exponential decay has the consequence that the Brillouin lines in the spontaneously backscattered spectrum must have a Lorentzian shape characterized by a full width at half maximum (FWHM or ΔνB) as shown in Figure 2. Typical linewidth value is 30 to 50 MHz dependending on the fiber type. The lorentzian distribution that models the product of the spontaneously scattered pump ligthwave is expressed as g (ν ) =

gB , 2 1 + (2 [ν − ν B ] Δν B )

(2)

where gB is the Brillouin gain coefficient. For single-mode fibers, gB varies from 1.12x10-11 to 5x10-11 m/W, depending on core doping and structure [6]. It is common that g be called the natural Brillouin gain of the fiber. An estimate of the scattered power can be obtained by assuming that in spontaneous regime, the number of Stokes photons generated is proportional to the number of acoustic phonons (Nph) present in the medium. At room temperature T, Nph≈kBT/hνB which is an approximation of Bose-Einstein distribution (h is Planck constant and kB is Boltzmann coefficient). The scattered power can then be evaluated by P = NhνsΔνB. For the typical values of Figure 5, the power is about 0.5 nW. Any temperature or mechanical stress would change the density of medium, and, in consequence, both n and VA. The temperature variation relates to the Brillouin frequency as [7]

ν B (T ) = ν B (Tr )[1 + CT (T − Tr )],

(3)

where T is the temperature and Tr is the reference temperature. A typical value for the temperature coefficient CΤ is 1.05 MHz/oC. Culverhouse demonstrated the feasibility of a temperature sensor based on the mechanism of Brillouin scattering [8] (Culverhouse 1989). When temperature remains constant, the Brillouin frequency shift νB(ε) relates to the applied strain as [7]

ν B (ε ) = ν B (0)[1 + Cε ε ],

(4)

where νB(0) is the unstrained Brillouin frequency shift and Cε is the strain proportionality coefficient. A typical value for Cε is 0.0550 MHz/με. b. Distributed measurement and sensor configuration

Distributed information is obtained by applying time, phase or frequency domain modulations of the measuring signals emitted by a laser. The distributed mode of operation can also be achieved by combining two or three of the modulation schemes. Two interrogator configurations can be implemented implying the use of two distinct regimes of Brillouin scattering which are known as spontaneous and stimulated Brillouin scattering (SBS).

96

F. Ravet

Spontaneous scattering which relies on the detection and the analysis of the backscattering of a modulated pump signal (Figure 2); depending on the type of modulation, the interrogator is commonly called Brillouin Optical Time Domain Reflectometer (BOTDR), Optical Frequency Domain Reflectometer (BOFDR), Brillouin Optical Coherency Domain Reflectometer (BOCDR) etc… Stimulated Brillouin Scattering (SBS) which relies on the detection and the analysis of a backscattered lightwave which is the product of the interaction of a pump and a probe signals (Figure 2); depending on the type of modulation, the interrogator is commonly called Brillouin Optical Time Domain Analyzer (BOTDA), Optical Frequency Domain Analyzer (BOFDA), Brillouin Optical Coherency Domain Analyzer (BOCDA) etc…

Fig. 2 Sensing configurations for spontaneous (a) and stimulated (b) Brillouin scattering. More specifically, BOTDR and BOTDA are presented here

BOTDR and BOTDA are the most commonly used configurations and are commercially available. They present inherent advantages in term of optoelectronics simplicity and measurement speed. The BOTDR requires access to only one fiber end. The launched lightwave is spontaneously scattered by the acoustic waves everywhere along the fiber length. The scattered light is then collected and analysed at the input end. Spatial information, and, hence, events location is given by measuring the round trip of a pulse propagating in the fiber. This configuration has been implemented by various research teams. In the BOTDA case, two lightwaves, the pump and the probe signals, are launched into the fiber in a counter-propagating configuration. The simultaneous presence of the Stokes and the pump waves generate a beat signal that reinforces the acoustic wave in the fiber when the beams frequency difference is equal to the Brillouin frequency. The coupling mechanism between the two lightwaves is electrostriction. The scattering of the pump is then enhanced, leading to its depletion and the input probe beam is amplified. The probe is also called Stokes as it corresponds to the frequency downshifted peak. The Brillouin spectrum can be recorded by tuning the frequency difference between pump and Stokes waves. One of the two lightwaves is pulsed and spatial information is obtained by the measuring the pulse round-trip [9]. Various parameters need to be considered when comparing BOTDR and BOTDA configurations. First, one has to keep in mind that the sensor must be


97

implemented on the field. It must then be simple to install and complete the sensing operation as quickly as possible. Second, some of the sensor performances are critical. Those are the spatial resolution, which indicates the smallest detectable event size, the frequency resolution, which is the smallest Brillouin frequency shift that can be measured, and the measurement range, which is the longest length overwhich the sensor can make an accurate data acquisition. The BOTDR has obvious advantages over the BOTDA. First, it only requires access to one fiber end. Second, if the fiber breaks, which is common in the field when the fiber is heavily stressed, measurements can still be done along the remaining section. On the drawbacks side, the BOTDR approach relies on the emitted spontaneous intensity. These levels are usually low, necessitating long averaging time to achieve a satisfactory signal-to-noise ratio (SNR). That is not the case with the stimulated configuration where the intensity at the detector is significantly larger, reducing the overall measurement time and allowing the use of shorter pulses. If we keep in mind that a fiber sees its attenuation increase when it is heavily stressed, the BOTDA will still be able to continue the monitoring operation over the whole fiber while the BOTDR is blind or offer poor performance behind the strained point. Significant fiber loss increase are common in SHM applications making the BOTDA their technique of choice. The use of SBS based interrogator makes possible the compensation for fiber loss increase and strain resolution degradation. A typical commercially available Brillouin interrogator system the DITEST which is an innovative high dynamic range laser-based monitoring system based on SBS [10]. The inherent stability of the system comes from the use of a single laser source and a high speed electro-optic modulator for the generation of both pump and probe signals. The intensity of both optical signals can be controlled in order to have the highest possible signal-to-noise ratio and reduce the acquisition time. The frequency difference between pump and probe signal is precisely controlled by the modulation frequency applied to the electro-optic modulator, leading to 10-5 precision on the frequency determination. Typically, the DITEST system performs strain profile measurement with a 10 με resolution and a spatial resolution better than 1 m over the first 10 km. For SHM applications, the interrogator is designed to handle short distance and large, optical budget (over 20 dB). 50’000 distance points can be measured with a minimum sampling interval of 0.1 m. The acquisition time (time to get one complete profile) may vary from a second to 10 minutes depending on the application requirements. c. Strategies for temperature compensation with the Brillouin sensor

In all laboratory tests, the temperature of the environment is controlled and variations are smaller than 1oC. Such situation is rarely encountered on the field where temperature variations are part of the measurement conditions. In fact, field implementation requires that the temperature influence on strain measurement must be compensated. Various strategies can be used involving the simultaneous monitoring of strain and temperature with the Brillouin sensor. The simplest consist in laying out an adjacent stress free fibre [11] or in gluing the jacket of a loose tube optical fibre cable. In that last case, all mechanical changes of the structure will only affect the cable while the fibre remains unstressed. This approach has led to

98

F. Ravet

the development of specialty cables which are specifically designed to measure temperature or strain [12]. It is also possible to measure simultaneously strain and temperature with a single fibre. Sensing can then be achieved by monitoring Brillouin frequency shift and Brillouin peak power[13]. Simultaneous temperature and strain can also be achieved by using fibres with multiple Brillouin peaks such as DSF, LEAF type [14] and PCF type fibers1 [15]. Finally, the distinct Brillouin behaviour of the slow and fast axis of polarisation maintaining fibers is another approach to measure strain and temperature at once [16].

3 Sensing System Operation and Model d. Performance definition i.

Spatial resolution

After Beller, “spatial resolution indicates instrument ability to resolve two adjacent events” (in Derickson 1998). In the case of the Brillouin sensor, the spatial resolution is the ability of the instrument to resolve two adjacent sections of distinct Brillouin frequencies, induced by either strain or temperature. The length over which the interaction between pump and Stokes wave occurs determines this parameter. This interaction length is the pulse length and it is defined as vgΔτ where vg is the pulse group velocity in the optical fiber and Δτ is the pulse width. The spatial resolution is then expressed as w = v g Δτ 2 .

(5)

The factor ½ comes from the backscattered nature of the detected signal. It accounts for the round trip of the signal in the fiber [18]. Since in silica fiber the group velocity is vg ≈ 2x108 m/s the rule of thumb that 10 ns corresponds to 1 m spatial resolution can be used. Based on this definition, a given temperature/strain that spreads uniformly over a distance greater than the spatial resolution is measured with the lowest measurement uncertainty. If a local temperature/strain change occurs in a distance scale smaller than the preset spatial resolution, it might still be detected but the change will not be measured with the lowest uncertainty. ii.

Brillouin frequency resolution

The Brillouin frequency resolution is defined as the smallest Brillouin frequency that can be resolved at a given location along the fiber. It is directly related to the noise of the measurement. The noise includes spontaneous, short duration deviations in output (reading) about the mean output (reading), which are not caused by strain or temperatures changes. Noise is determined as the standard deviation about the mean and is expressed in frequency units. The measurement precision is a measure of the agreement between repeated measurements of the same property under identical, or substantially similar, conditions; the precision is defined. The frequency resolution is defined as twice the standard deviation of the noise (+/twice the standard deviation includes 95.4% of the measurements). Once the fiber 1

Dispersion Shifted Fiber, Large Effective Area Fiber, Photonics Crystal Fiber.


99

sensor is calibrated, it is always possible to display the measurement resolution in term of strain or temperature [19]. Horiguchi proposed a relation that gives an estimate of the minimum detectable frequency change [20]. The relation accounts for the measurement dependence to noise and Brillouin spectrum characteristics and is expressed as δν B =

Δν B 2SNR

,

(6)

where SNRo is the signal-to-noise power ratios. As it appears in Eq.(6) Brillouin frequency resolution not only depends on the signal level but also on the spectrum width. Common distributed Brillouin sensors are then limited in frequency resolution when the pulse spectrum is larger or equal to ΔνΒ which is the case when pulse width is shortened to achieve better spatial resolution [21]. For most of the Brillouin sensor configurations, there is clearly a trade-off between spatial and frequency resolution [22]. Both statistical and analytical definitions specifically apply to the minimum detectable change when the fiber Brillouin frequency is uniform over spatial resolution. It does not address the issue of the frequency resolution when νB is not uniform over w. A definition of the spatial resolution accounting for non-uniform νB distribution is discussed in detail in section 4.b.ii of the present work [23]. iii.

Brillouin frequency accuracy

In general the term accuracy is a generic qualitative word. It should be associated with “uncertainty” or “calibration uncertainty”. The measurement uncertainty depends on the calibration precision, i.e. on the quality of the calibration setup and procedure. For instance, the calibration of a piece of fiber as a temperature sensor requires that a traceable reference temperature sensor with given precision be available. It is therefore impossible to mention the instrument calibration uncertainty without including a given fiber sensor, that has been calibrated with a given precision [19]. iv.

Dynamic range

The dynamic range (DR) expressed in dB is defined as the difference between the maximum input power to the photodetector and the smallest optical power level that can be detected. It is a measure of the total loss that can be accommodated by the instrument when performing a measurement [24]. The dynamic range is then a function of the power launched into the fiber, the Brillouin interaction, the loss of the components and the receiver characteristics [20]. It can be expressed as DR =

{

1 1 1 ⎫ Pmax − G p − α comp − Pd − SNR req + SNR imp ⎬, 2 2 2 ⎭

(7)

where Pmax is the maximum fiber input power, Gp is the Brillouin loss, αcomp is the loss of any component located along the fiber, SNRreq is the signal-to-noise ratio required to achieve a target frequency resolution (Eq.(6)), and, SNRimp is the signal-to-noise ratio improvement obtained by electrical signal processing (averaging, electrical amplifiers). Various causes limit the maximum fiber input power: 1)

100

F. Ravet

Pmax must be smaller than the SBS threshold for BOTDA configuration [25], [26]; 2) Pmax has to be lower than the maximum power that the receiver can handle; 3) another technological constraint on Pmax is determined by the available power at the laser output. The factor ½ in front of the bracket is introduced to account for the fact that the signal suffers twice the loss as it propagates in both directions before being measured [20], [24]. The other ½ factors account for the fact that these SNR’s are expressed in electrical decibels.

4 System Operation and Model The sensing configuration chosen is a BOTDA as illustrated in Figure 2 in which the pump is a cw and the probe is pulsed. From performance definition section it is obvious that temperature and strain accuracies are affected by the Brillouin spectrum shape. The Brillouin spectrum shape is influenced by the pulsewidth (Δτ ), shape, pulse power (P ) and pump cw power (Pcw), fiber length (L), and ON–OFF ratio of the optical pulse (or extinction ratio ER). The pulse power is always composed of an AC (Ppk) and a DC (Pb) component due to its finite ER as illustrated in Figure 3 due to the electro-optic modulator bias. pulse, stokes

Ipk P

Pulse

ER Base

Ib P

t cw, pump

t Fig. 3. Model of pump-stokes (pulse) interaction.

The interaction between probe and the pump lightwaves can be modelled by the steady-state-coupled intensity2 equations [1], [5], which are d I cw = − gI p I cw − αI cw , dz

2

(8)

We need to introduce the relationship between the optical power in the fibre, P, and the total intensity distributed in the fibre cross-section. Assuming Gaussian radial distribution the Power=I0Aeff. As it appears in this relation, when the intensity has a Gaussian distribution, the power is the product of the peak intensity Io with the effective area of fibre Aeff. The effective area can be interpreted as the area of the fibre cross section over which the peak of the intensity distribution is constant [5].


d I p = − gI cw I p + α I p . dz

101

(9)

where Ip is the Stokes intensity, which is pulsed; Icwis the pump intensity; z is the position along the fiber in which the Stokes pulse is launched from z = 0and the pump from z = L; α is the fiber natural attenuation; and g is the natural Brillouin gain of the fiber. Rigorously, these equations can only be used to describe steady state or long pulses (Δτ > 10 ns) interaction. We also assume that the gain coefficient g depends on position. The position dependence of g is attributed to the fact that fibers do not have a perfectly uniform Brillouin frequency distribution. In addition, the purpose of a sensor is to detect Brillouin frequency variations induced by the environment. The coupled intensity equations can be solved under the weakly depleted approximation which can be applied to most of the sensing cases. Let us give two major reasons. First, the pulse length, and hence the Brillouin interaction length, exceeds rarely 100 m, in which case the pump depletion is always weak. Second, the DC component of the pulse can significantly deplete the pump if the fiber is longer than a few hundred meters. We then want to make sure that Pb is smaller than the SBS threshold for BOTDA configurations [25], [26]. Hopefully, it is in practice a fraction of the pulse peak power (e.g. a few tenth of mW at the very maximum). We make the assumption that pump is weakly depleted. If we choose to let the pulse enter into the fiber at z = 0 and to launch the pump from z = L, then initial conditions to Eqs.(8) and (9) are IP0 and IcwL, for the Stokes and the cw pump lightwaves respectively. Solving Eqs. (9) leads to the following expression I P (ν , z ) = I P 0 exp[g (ν , z )I cwL

e −αL

α

(e

αz

⎤ − 1 − αz ⎥. ⎦

)

(10)

We now follow the assumption that the Stokes wave is constituted of two components [27], [28], [29]: one is time dependent (AC part of the pulse), which is characterized by the pulsewidth (Δτ) and the peak intensity (Ipk); the other is the base (DC part) which is a continuous wave signal (cw) and determined by its intensity (Ib). Figure 3 describes this model. The Stokes intensity is then the sum of these two components IP=Ipk+Ib. Solving Eq.(9) with these intensities, and their spatial dependence expressed as Eq.(10), we obtain the total Brillouin loss, which can be expressed in the following form for an arbitrary length l [31]

{

GT (ν , z ) = exp αw − ∫

z +l

z

g (ν , ζ )(I pk + I b ) ⎡ ⎤ ⎫ e−αL αζ exp ⎢ g (ν , ζ )I cwL e − 1 − αζ ⎥ dζ ⎬ α ⎣ ⎦ ⎭

(

)

(11)

We see that the AC and DC components of the pulse can be separated leading to the following generic form of the Brillouin loss GT (ν , z ) = G pk (ν , z )Gb (ν , z ),

(12)

102

F. Ravet

where Gpk is the Pump-AC Brillouin loss and Gb is the Pump-DC Brillouin loss. Due to their very distinct natures, we want to separate the calculation of these two interactions. First, we consider the pump-AC interaction. The steady state approximation is valid for pulses larger than the phonon lifetime (Δτ > 10ns), which is equivalent to a spatial resolution w>1m. We know that nanosecond pulses provoke the broadening of the Brillouin loss spectrum and then reduce the strain (temperature) measurement accuracy, while such pulses are required to achieve centimetre spatial resolution. As the broadening is induced by the wide spectrum of the pulse, we assume that each pulse spectrum component excites individually the phonon field whose spectrum is given by a Lorentzian distribution. These individual excitations are adding up over the pulse spectrum frequency range. That mechanism is described mathematically by a convolution product. Therefore, we account for the pulse effect in these equations by replacing g with the transient gain coefficient gpk, defined as the convolution of the Brillouin natural gain g(ν,z) with the distribution Ppkb(ν) of the power spectrum of the pulse [22], [30]. g pk (ν , z ) = g (ν , z ) * b(ν ).

(13)

If the pulse waveform has a rectangular shape, we can then calculate the convolution analytically [31]. Finally, we integrate equation II.45.b for the AC part of the Stokes wave at any position over the spatial resolution (w) [31]and obtain the Brillouin loss spectrum Gpk at z:

{

G pk (ν , z ) = exp αw − ∫

z+w z

g pk (ν , ζ )I pk 0 ⎤ ⎫ ⎡ e −αL αζ exp ⎢ g pk (ν , ζ )I cwL e − 1 − αζ ⎥ dζ ⎬, α ⎦ ⎭ ⎣

(

)

(14)

Now, we want to evaluate the interaction between the base and the pump wave. Due to the DC nature of the base, we assume that its interaction with the pump can be modelled by the steady state equations without additional assumptions. Here the Stokes intensity reduces to Ib=ERIpk where ER is the extinction ratio of the optical pulse. The integration interval extends to the whole fiber length. The Brillouin loss spectrum contribution from the base-pump is then expressed as

{

Gb (ν , z ) = exp α ( z − L ) − ER ∫ g (ν , ζ )I pk 0 L

z

⎡ ⎤ ⎫ e −αL αζ exp ⎢ g (ν , ζ )I cwL e − 1 − αζ ⎥ dζ ⎬. α ⎣ ⎦ ⎭

(

)

(15)

When it is assumed that the Stokes pulse is the sum of two parts, DC and AC intensities, that are decoupled but interacting individually with the pump, the total Brillouin loss spectrum is the product of Gpk and Gb. GT of a uniform Brillouin frequency distribution is characterized by a single peak spectrum whose FWHM, Γ, is close to ΔνB when pump is weakly depleted and under steady state condition. In general, Γ varies with L, z, Ipk0, IcwL and ER, which are the cause of spectrum distortion.


103

e. Model discussion – Unstressed fiber

We conducted a systematic study of the influence of pulse width, extinction ratio on the linewidth of the Brillouin loss. Brillouin frequency was assumed to be uniform along the fibers. Typical results for short and long fibers are presented in Figure 4. This agreement with exact solutions strengthens our confidence in the validity of the present approach. As shown in Figure 4, Γ increases when the pulse width decreases until it reaches a maximum. Then Γ drops to a value close to ΔνB. Small ER clearly limits the increase of the linewidth. When Δτ ≈ 10 ns, the observed broadening must be attributed to the pulse spectrum widening. For nanosecond pulses, Γ drops due to the DC level dominating the Brillouin interaction over long fiber length and relative small contribution of the pulse portion. If large pulses are used (Δτ >> 10 ns), the pulse spectrum is much narrower than ΔνB. The Brillouin loss spectrum tends to be similar to the spectrum resulting from the interaction of cw pump and Stokes lightwaves. The main difference between short and long fibers, respectively Figure 5.(a) and (b), must be attributed to increased influence of the base. Pump and base interaction extending over the whole fiber length, Γ tends to remain close to steady state value when the fiber is long as it appears in Figure 5.(b). Γ also varies with position as shown in Figure 6(for a 2500 m long fiber and w = 1 m. In the ER=25dB case, it is observed that Γ first decreases. In the last kilometres, Γ starts to rise significantly. The increase of Γ is very weak when considering the ER = 10 dB situation. Similarly to Figure Figure 5, these curves confirm the role played by the base in mitigating the spectrum broadening. The effect of spectrum narrowing with increasing distance can be attributed to the enhancement of the scattering of Stokes spectral components near the peak of the resonance while the pulse is propagating along the fiber. Equivalently, a broader spectrum at the pulse input can be attributed to the effect of pump depletion. Spectrum widening observed at the fiber end can be understood by analysing Figure 5.(b) representing the base (Gb) contribution to the total Brillouin loss (GT) as a function of position and ER. In the case of high spatial resolution sensor (w ≤ 1 m), the change of Gpk as a function of z is small relative to Gb and does not depend on ER. Close to the pump input, the Brillouin loss is smaller (Gb → 1 when z → L). That effect is emphasized and affects Gb over the whole fiber length (Gb≈1) when the pulse extinction ratio is large. It is then possible to observe the pulse spectrum broadening in the last few hundred meters of the fiber. Spectrum narrowing with position can be experimentally evidenced as illustrated in Figure 6. The experiment was carried out with a Brillouin sensor using two DFB laser in the 1.55 μm region. The typical linewidth of these lasers is about 10 MHz and contributes to broaden the Brillouin loss spectrum. Under weak depletion condition, we measured that Γ ≈ ΔνB ≈ 45 MHz. The phenomenological model is used to calculate the spectrum along the fiber. The FWHM of the reconstructed profile is plotted on Figure 6. It is obvious that the reconstruction matches the experimental trend, which is a confirmation that the model is capable to simulate the effect of moderate pump depletion that influences the spectrum shape.

104

F. Ravet 150

FWHM [MHz]

130

35

13dB 15dB 20dB

(a)

15dB

34

Γ [MHz]

110

Γ [MHz]

13dB

(b)

90 70

20dB

33

32

50 30

31

0

20

40 60 Pulse ΔτWidth [ns][ns]

80

100

0

20

40

60

Δτ [ns]

80

100

Fig. 4 Brillouin loss spectrum width as a function of pulsewidth and extinction ratio: (a) Pp0 =10 mW, PcwL =5 mW, L = 10 m, z = 5 m; (b) Pp0 = 10 mW, PcwL = 5 mW, L = 10000 m, z = 5000 m [33] 1.0 10dB (a) 25dB

(b)

0.9

Gb

Γ [MHz]

60

40

0.8 0.7

20 0

500

1000 1500 2000 2500

z [m]

0

500

1000 1500 2000 2500

z [m]

Fig. 5 (a) Brillouin loss spectrum width as a function of position and extinction ratio (filled symbols:ER = 10 dB; open symbols: ER = 25 dB); (b) base contribution to total Brillouin loss as a function of position and extinction ratio (filled symbols: ER = 10dB; open symbols: ER = 25dB), these values are calculated at the Brillouin frequency; Simulation parameters are: ν = νB, Pp0 = 25 mW, PcwL = 5 mW, Δτ = 10 ns, L = 2500 m [33]

Fig. 6 Measured (light grey line) and calculated (dark line) Brillouin loss spectrum width, Γ, as a function of position for a uniform fiber; experimental and simulation parameters are Pp0 = 20 mW, PcwL = 5 mW, L = 1800m, Δτ = 10 ns, ER = 10 dB, ΔνB = 45 MHz [33]


105

f. Model discussion – Stressed fiber

Brillouin Frequency Shift

To study the effect of changing stress length and strength on the spectrum shape, we assume that the sensing fiber has a uniform Brillouin frequency (νB) over its whole length (L) except for a short section at distance z whose spatial extension δl is smaller than the spatial resolution (δl ≤ w) and has a different Brillouin frequency shift (νBs). As appearing in Figure 7, within the spatial resolution at z, the Brillouin frequency shift (νBs) is constant over δl, while the rest of the pulse covers w- δl with a Brillouin frequency shift of νB. We use the phenomenological model for w > 1 m to generate composite spectra for various combinations of νB, νBs and δl. Smaller spatial resolution could be considered but the analysis would be more complex without bringing useful information. We introduce the normalized Brillouin frequency shift ΩBs=(νBs-νB)/Γ where Γ is the FWHM of the normalized Brillouin loss spectrum for given sensing parameters. νBs νBs, ΔνB νB, ΔνB νB

Position z+δl

z

z+w

Normalised Brillouin Loss

Fig. 7 Brillouin Frequency distribution within the length of the spatial resolution. Both sections have the same Brillouin linewidth but have a distinct Brillouin frequency [34]

1.0

0.8

a 0.6

b

c

0.4

0.2

0.0 12750

12775

12800

12825

12850

12875

12900

ν (MHz)

Fig. 8 Composite spectra for three distinct Brillouin frequency shift. Spectra a, b, and c are respectively associated with a Brillouin frequency of 12810MHz, 12820MHz and 12860MHz. The unstressed Brillouin Frequency shift is 12800MHz. The simulation parameters are Ppk0 = 30 mW, PcwL = 5 mW, L = 1000 m, z = 500 m, w = 20 m, δl= 5 m [34]

106

F. Ravet

Figure 8 illustrates how the Brillouin spectrum shape changes when νB, νBs and δl are varied. In this case, we have that δl = 5 m, νB = 12800 MHz while νBs is 12810MHz (curve a), 12820MHz (curve b) and 12860MHz (curve c). Curve (a) is a single-peak distribution but lightly distorted. Curve (b) is still a single-peak distribution but it appears to be skewed. The central frequencies in Curves (a) and (b) are shifted from stress-free frequency (12800MHz), even if δl (stress section) is only 25% of the spatial resolution. The curve (c) has two peaks at 12800MHz and 12860MHz respectively due to the loose part of the fiber and the stressed section. The curves in Figure 8 clearly show the influence of the both stress length and strength on the spectrum shape. There is a transition from a single peak spectrum to a dual peak spectrum, which is depending on the stressed length. i. Length-Strength diagram

The spectrum shape is driven by the combination of the normalised Brillouin frequency shift and the stressed section length. Such behaviour is summarised in a Length-Strength diagram represented in Figure 9. This diagram is the result of a systematic analysis of the spectrum shape as a function of the normalised Brillouin frequency shift and the stressed section length. We search for the number of peaks present in the spectrum for each combination of δl/w and ΩBs. We record each couple (δl/w,ΩBs) corresponding to the transition from single to dual peak spectrum. The values are then reported on a single diagram (Figure V.4) in the form of a curve that shows ΩBs as a function of δl/w. Below the curve, only one peak can be seen; above the curve, two peaks are present. When the normalized Brillouin frequency shift is below 0.65, only one peak is observed whatever the value of

2

ΩBs

1.5 Two peaks regime

1 One Peak Regime, Major contribution from unstressed peak

One Peak Regime, Major contribution from stressed peak

0.5 0

0.25

0.5

0.75

1

δl/w Fig. 9 Length-Strength diagram: this figure reports the normalized Brillouin frequency shift at which two peaks start to be observed as a function of δl/w. Below the curve, only one peak can be seen. Above the curve, two peaks are present. The simulation parameters are Ppk0 = 30 mW, PcwL = 5 mW, L = 1000 m, z = 500 m, w = 20 m [34].


107

δl/w. It is clear that there is a threshold frequency Ωth that governs the number of peaks in the spectrum and that is determined by the Length-Strength diagram (LS). ii. Rayleigh Equivalent Criterion


There are two frequency regimes characterized by single and two-peaks Brillouin loss spectra. A frequency threshold Ωth sets the border between the two regions. As long as ΩBs is very different from Ωth, there is no ambiguity in finding peak frequency. This claim is not valid when ΩBs ≈ Ωth, because the discrimination of the two peaks requires the peak search to be very precise. Besides, experimental data are contaminated with noise. That makes it more difficult to distinguish the two regimes. We want to introduce a practical and reproducible criterion that allows the unambiguous detection of multiple peaks and to determine the smallest resolvable frequency shift (Ωres) using Rayleigh criterion3 [37]. An equivalent criterion that applies to Brillouin Spectrum can be derived by making the assumption that both the normalised Brillouin spectrum and I must have the same FWHM [35]. If we define the normalised frequency in I to be β=υπν/Γ, it is easy to find that υ =1.7718 for Brillouin spectra that are Lorentzian like. Figure 10 shows Brillouin loss spectrum obtained by simulation of the phenomenological model. The dip amplitude (minimum of the Brillouin loss spectrum comprised between the two peaks) is 0.75 corresponding to Ωobs = Ωres = 1.13. We propose to define the dip amplitude as the Rayleigh Equivalent Criterion (REC).

1.00

Ωres=1.13

0.75 0.50 REC=0.75

0.25 0.00 -2.0

-1.0

0.0

1.0

2.0

3.0

Ω

Fig. 10 Definition of the Rayleigh Equivalent Criterion for simulated Brillouin loss spectrum with the following parameters L=1000m, z=0, Pp0=30mW, PcwL=5mW, w=20m [34][35]

3

Rayleigh criterion is known to be a criterion that allows the determination of the smallest resolvable frequency difference. It applies to two distributions of equal intensity whose equations have the generic shape I = sinc2(β) where β is the normalized frequency. The criterion assumes that these two peaks can be resolved as soon as the maximum intensity of the first peak coincides with the first minimum of the second peak, which happens for β = π. The distance between these two peaks is then the smallest resolvable frequency difference. The minimum between the two peaks has an intensity of 8/π2.

108

F. Ravet


1.2

A 1.0

REC B

0.8

0.6

C

0.4

0.2

0.0 12700

12750

12800

12850

ν (MHz)

12900

12950

13000

Fig. 11 Experimental Brillouin loss spectra measured at 20m at the boundary between the strained and unstrained sections for three strain values (A: 510με; B: 638 με; C: 893 με) [35]

Figure 11 illustrates how the REC can be used in measuring stresses that are shorter than the spatial resolution. These spectra come from the same experiment that illustrates the REC dependence against the normalised Brillouin frequency shift. The discussion focuses on the location at the boundary between the strained and unstrained sections. Profiles A (510 με) and B (638 με) show that the pulse covers more than one stress. According to the REC they are not distinguishable. Note that spectrum B clearly experiences two peaks and then the stress could be measured but an uncertainty much larger than 5% must be expected. The same criterion states that Spectrum C (893 με) has a stressed section that can be clearly identified and measured, with an error much lower than 1%. The stressed contribution frequency is estimated to be 12873.33 MHz when the peak frequency is measured. The frequency shift associated with a strain of 893 με is actually12878.90 MHz. Also the location accuracy has been reduced to ½ of the pulse length. iii.

Form Factors

When two peaks can not be resolved according to the REC (i.e. the spectrum lies in the single peak region of the LS diagram), further analysis of the spectrum shape can be carried out and then useful information deduced. That is possible by introducing two form factors, asymmetric (FA) and broadening (FB) Factors that describe the distortion of single peak spectrum [36]. The form factors are defined as FA =

Γ+ Γ−

FB =

Γs Γ

(16) (17)

where FA is the asymmetric factor and FB the broadening factors, Γ+ is the half width at half maximum of the right side of the spectrum, Γ− is the half width at


109

half maximum of the left side of the spectrum and Γs is the FWHM of the stressed fiber spectrum. Brillouin spectra measured at each location are analysed to extract these three parameters as illustrated in Figure 12.

Relative Brillouin Loss

1.2 1.0 0.8 0.6 0.4

Γ−

Γ+

0.2 0.0 12750

Γs = Γ− + Γ+ 12780

12810

12840

ν [MHz]

12870

12900

12930

Fig. 12 Definition of the width parameters on an experimental Brillouin loss spectrum of a strained section of a single-mode optical fiber [36]

The asymmetric parameter, FA, indicates the presence of large but short strain components. FB describes the broadening of the Brillouin loss spectrum induced by non-uniform strain distribution. Let us discuss various strain regimes associated with the form factors value. Figure 13shows various cases with the same peak frequency but different strain distributions. When FA =1 and FB =1, the strain distribution is uniform. The spectrum is simply shifted and a peak finding approach is enough to characterize the status of the structure (Figure 13.(a)). If the spectrum is still symmetric (FA =1) but FB >1, then the distribution is non-uniform [38]. Peak finding technique describes the global behaviour of the structure but it fails to detect the presence of strain over section shorter than the pulse-width. The strain distribution becomes asymmetric when FA ≠1 as shown in Figure 13.(b), which is an indication that the strain distribution is non-linear [39]. For FA >1, the strain distribution is non-uniform over the pulse length: a short length strain component, whose amplitude is large, and, a long strain component, whose amplitude is small, are covered by the pulse. In other words, the spectrum presents a broadening happening on the right of the peak frequency. It indicates that small defects inducing large strain start to build up in the structure. In the case of FA 1

FA=1, FB=1

0.4 0.2 0.0 -2.5

-1.5

-0.5

0.5

1.5

Normalised Frequency

2.5

1.2

(b) 1.0 0.8

FA>1, FB>1

FA1

0.6 0.4 0.2 0.0 -2.5

-1.5

-0.5

0.5

1.5

2.5

Normalised Frequency

Fig. 13 Normalised Brillouin loss spectrum for various strain profiles included within w: (a) uniform strain (FA=1, FB=1), linear strain (FA=1, FB>1); (b) non-linear strain with short components whose amplitudes are larger than the main strain contribution (FA>1, FB>1), non-linear strain with short components whose amplitudes are smaller than the main strain contribution (FA1) [36]

5 Data Analysis Methodology g. Proposed methodology

The model introduced previously not only explains the physics behind the instrumentation. It can also be used in a spectrum reconstruction scheme where the spatial distribution of νB is unknown which is the most common situation encountered on the field. The useful information is extracted from a careful analysis of the spectrum shape using the tools introduced in previous section. The first step would consist in determining to what part of the LS diagram the spectrum belongs. As mentioned, two types of spectrum shape are observed depending on the Brillouin frequency difference between stressed and unstressed components, as well as stressed section length. One type of spectrum has a single distorted peak which is broadened and asymmetric in most of the cases. The other type of spectrum has two peaks. Because the spectrum type is a function of ΩBs and δl/w, the classification in these two categories according to their dependence in stress value and length can be done as illustrated in Figure 14 by the LS diagram. There is a clear border (shown as a line in Figure 14.(a)) between the region where a single distorted peak is present (Figure 14.(b)) and the region where two peaks can be identified (Figure 14.(c)). The profile is then influenced by stress length and amplitude. When two peaks are clearly identified, as shown on the spectrum of Figure 14.(c), stressed section length and amplitude are translated respectively in relative peak height, γ, and peak frequency. The relationship between γ, ΩBs and δl/w is shown in Figure 15. Handling multiple integrations at each position, as the phenomenological model would require, is not necessary. Instead we propose to calculate GT , assuming that g is constant over w but composite i.e. g at z is a linear combination of Lorentzian shape distributions defined as

2.0 (a)

Double peak region

1.6 (c)

ΩBS

1.2 0.8 (b) Single peak region

0.4 0.0 0.0

0.1

0.2

δl/w

0.3

0.4

0.5

Relative Brillouin Loss Relative Brillouin Loss


111

1.0

(c)

0.8 0.6 0.4 0.2 0.0 1.012600 0.8

12800

ν [MHz]

13000

(b)

0.5 0.3 0.0 12600 12700 12800 12900 13000

ν [MHz]

Fig. 14 (a) Length-Stress diagram for non-uniform Brillouin frequency over pulse length: the continuous line is the border between (b) single peak spectrum (δl/w = 0.28, ΩBs = 0.72) and (c) two peaks spectrum (δl/w = 0.40, ΩBs = 1.60). Simulation parameters are Ppk0 = 10 mW, PcwL = 8 mW, L = 100 m, w = 1 m, ER < 20 dB [33].

g (ν , z ) =

γ i ( z )g B 1 N (z) , ∑ N ( z ) i =1 ⎛ ν ( z ) − ν ⎞ 2 ⎟⎟ + 1 ⎜⎜ 2 i Δν B ⎠ ⎝

(18)

where γi is the height and νi the peak frequency of the ith peak in spectrum, N is the number of peaks. As it appears in Eq.(17), γi, νi and N depend on the location in the fiber. The combination of these parameters is unique as it is based on the analysis of the spectrum shape, as it will appear in the following sections. When the Brillouin spectrum is composed of multiple peaks, the present signal processing approach does not require to know the exact relationship between γ and δl/w as shown in Figure 15. Here we would implement a peak search routine and associate each detected maximum with a pair (γi,νi). Near the borderline as defined in Figure 14, the distinction between single and double peak is not always unambiguous, particularly when the experimental data are contaminated with noise. To overcome that difficulty we introduce the Rayleigh Equivalent Criterion (REC). It states that two peaks are resolved if the minimum between two apparent maxima is lower than 75% of the lowest of the two peaks. We will use this criterion to separate the single peak region from the double peak region without ambiguity. In the two peaks region, the identification of the various components is easier and the pair (γi,νi), (i=1,2), are determined without ambiguity, being the height and the frequency of the detected peak. When the spectrum only experiences one but distorted peak as shown in Figure 14.(b), the estimation of (γi,νi), (i=1,2), is more complicated. It is clear that in that case, the spectrum appears broadened and

112

F. Ravet

1.00

ΩBs =0.79

γ

0.75

ΩBs=1.23

0.50 0.25

ΩBs =1.95

0.00 0.0

0.1

0.2

0.3

δl/w

0.4

0.5

Fig. 15 Relative peak height of a stressed section as a function of the normalized stressed section length and Brillouin frequency shift. The solid line is the peak height when two peaks can be observed in the spectrum (all these points are in the double peak region of the LS diagram). The dotted line is a linear extrapolation of the stressed section peak height when the stressed contribution is buried in the single peak spectrum (all these points are in the single peak region of the LS diagram). Simulation parameters: Ppk0 = 10 mW, PcwL = 8 mW, L=100m, w=1m, ER>3 Stationary

Constant, Large

FRP cracks

Observation

Non-uniform, large strain dominates

Reduced nonuniformity, large strain dominates

=20 ms), this was an important factor in system selection.

Fig. 5 System design

The Easyheat consists of the main induction heating control box which supplies power to the work head. The work head contains a transformer-coupled resonant circuit, including two capacitors and the excitation coil itself. The excitation frequency is dictated by the value of the capacitors, the inductance of the coil and the load on the circuit, ie. the material, volume and proximity of the sample under inspection. Preliminary tests with a variety of cameras showed that many of the interesting features of the measured temperature change in a material under inspection are within the first few tens of milliseconds of the heating period and, after numerous tests, it was decided that the minimum heating period would be set at 20 ms. Hence a fast frame rate was identified as a critical factor in camera selection, along with excellent thermal sensitivity. After auditioning several cameras for the

212

G.Y. Tian et al

system, the FLIR SC7500 was chosen for the work. The SC7500 is a Stirling cooled camera with a 320 x 256 array of 1.5 - 5 μm InSb detectors. The camera has a sensitivity of n2

coating

Fig. 1 Structure of an optical fiber

Early in the development of optical fiber technology it was recognized that apart from being an excellent transmission medium for communications due to its low attenuation and large bandwidth, optical fibers could also serve an important role for sensing. In this application the optical fiber is used to transport light to and from an optical transducer that modulates or changes some parameter of light (intensity, spectral content, polarization, phase, etc) in response to some measurand. The transducer can be the optical fiber itself, leading to an intrinsic fiber optic sensor, or some external device, which results in an extrinsic or hybrid sensor [1]. Optical fiber sensors (OFS) are very interesting for structural health monitoring (SHM) applications due to the advantages of this technology compared to conventional sensors. Some of these advantages are inherent to the physical and material characteristics of the optical fibers that made the sensors. For instance, OFS are immune to EM interferences because light propagates in optical fibers within a dielectric material in which interaction with electromagnetic fields can be neglected. This is certainly an asset in many applications where electric motors, generators or transformers are deployed. In addition, the dielectric characteristic of fibers makes them lightning-safe in applications such as wind power generators where the blades must be free from conducting materials. The fact that the transmission of light is confined to the fiber and the absence of electrical currents also makes optical fiber sensors intrinsically safe for deployment in explosive environments. Moreover, the optical fibers are typically made of inert material (mainly silica) resistant to most chemicals and to weathering effects and corrosion. Another important advantage of optical fibers for SHM is their small size with external diameters as small as 125 microns for widely deployed fiber types. This is very important when embedding OFS in order to make sure that the mechanical properties of the host material are not significantly compromised due to the introduction of the fiber [2]. There are also advantages related to the transmission characteristics of the fiber. For example, the light traveling in an optical suffer suffers very low attenuation of its power and this means that large structures can be covered with complex mesh routing of fiber without the need for regeneration of the signal in any active components. The other significant advantage of optical fiber for transmission is their extremely large spectral bandwidth, which leads to the possibility of

Optical Fiber Sensors for Structural Health Monitoring

337

propagating light of different wavelengths or colors simultaneously on the same fiber. Each wavelength can be used for the interrogation of a specific sensor in a wavelength-division multiplexing (WDM) scenario. Moreover, at each wavelength light can be pulsed very fast, leading to time-division multiplexing (TDM) schemes. Other possibilities of multiplexing OFS are also available such as coherence-division multiplexing or polarization multiplexing. Other advantages of OFS are linked to the implementation of the technology itself. There are many situations in which deployment of OFS leads to improved measurements over conventional sensors with simpler installation and maintenance. Moreover, OFS have demonstrated enhanced stability and reliability over time due to some of the characteristics previously discussed. Some OFS can be regarded as direct replacement for conventional sensor alternatives, providing similar measurements, but with additional performance that justifies their use. On the contrary, other types of OFS offer completely novel sensing possibilities that are simply not available with conventional sensors. An example is distributed sensors, which provide the magnitude of certain parameter, e.g. strain or temperature, at every position along an optical fiber. Distributed sensors for structural health monitoring are treated latter in this chapter. Finally, SHM applications can benefit from the extensive research that has been performed over the years to develop OFS for a very large variety of mechanical, physical and chemical parameters. Nevertheless, the most important measurands for this application are mechanical parameters such as strain, deformation or displacement and some physical parameters such as temperature. Other parameters of interest can be readily derived from these measurands by proper sensor design. For instance, the corrosion of reinforcing steel in concrete structures can be detected by the strain and deformation brought by the large radial pressure exerted on the surrounding material. In this chapter, we review the most significant OFS for SHM applications. The analysis focuses on the most important sensor types from the point of view of current commercial availability and potential for short-term widespread deployment. Mature, well-established technologies such as fiber Bragg grating (FBG) and interferometric sensors are highlighted first in sections 2 and 3. Then, sections 4 and 5 describe newer distributed sensor technologies based on Brillouin and Rayleigh scattering effects, which are currently having a very significant impact on SHM owing to the enhanced monitoring features that they provide.

2 Fiber Bragg Gratings Sensors One of the most commercially relevant FOS types is the fiber Bragg grating (FBG), which is finding widespread use as point temperature and strain sensor and a direct substitute for electronic counterparts such as strain gauges. FBG are made by inducing a periodic modulation of the refractive index in the core of an optical fiber. This modulation of the refractive index is made possible by the

338

A. Loayssa

photo-sensitivity effect, a nonlinear phenomenon by which the refractive index in the core of a fiber is permanently modified by exposure to ultraviolet radiation. The change Δn of the refractive index depends on the radiation dose, intensity and wavelength. There several techniques that can be used to write gratings on optical fibers such as interferometric methods and phase-mask techniques [3].

broadband light

Λ cladding core

λ Bragg grating

λB

λB

λ

transmitted light

λ

reflected light

Fig. 2 Uniform fiber Bragg grating as an optical filter for transmission and reflection of light

The simplest and more widely deployed FBG for sensor applications is the uniform FBG (UFBG). This grating can also be regarded as the basic building block of more complex types. In UFBG the modulation of the refractive index has a constant period Λ along the grating length. As schematically depicted in Fig. 2, this device behaves as a wavelength-dependent optical filter for transmission and reflection of light. It is found that overall light around the so-called Bragg grating wavelength λB is reflected when the following condition is satisfied:

λB = 2neff Λ

(1)

with neff the effective refractive index of the propagating fiber mode. According to basic optics phenomena, part of the light that propagates along the grating is reflected at each refractive index transition. Then, the condition in eq. (1) just conveys the fact that the multiple reflections of light of wavelength λB from each grating plane interfere constructively because there is relative 2π phase-shift in their propagation constants. Conversely, if the Bragg condition is not satisfied, the reflected light from each of the subsequent planes becomes progressively out of phase and will eventually cancel out. Only for light around λB the weak reflection from each grating plane reinforces itself and leads to a strong reflection. Furthermore, the light that is reflected is removed from the transmitted spectra. Therefore, overall a UFBG behaves as a narrow bandpass filter in reflection and notch filter in transmission. The transmission and reflection spectra can be measured, for instance, using a broadband incident light source or a narrow-band tunable laser. Other types of gratings that are less important for sensor applications are:


339

long-period gratings, in which light couples into cladding modes, chirped FBG, in which the period of the grating is varied along the length of the grating, or tilted FBG, in which the gratings are written at an angle to the fiber axis. Sensor applications of UFBG are based on the sensitivity of the Bragg wavelength to strain and temperature. From Eq. (1) the change in λB as a function of temperature or strain in the FBG can be derived as [3]:

⎛ ∂neff ⎛ ∂n ∂Λ ⎞ ∂Λ ⎞ ⎟⎟ Δl + 2⎜⎜ Λ eff + neff ⎟ ΔT ΔλB = 2⎜⎜ Λ + neff ∂ l ∂ l ∂ T ∂ T ⎟⎠ ⎝ ⎠ ⎝

(2)

where Δl is the change in grating length due to strain. The first term in the expression is related to the strain-optic effect by which a change in strain brings about a variation in the refractive index in the fiber. The second term is simply the grating period change which is proportional to the strain. The third term comes from the thermo-optic effect related to the dependence of refractive index on temperature. Finally, the last term comes from the thermal expansion coefficient of the fiber. The net result of all these variations is that the Bragg wavelength, i.e., the central wavelength of the narrowband filter, depends linearly on strain and temperature over typical ranges of interest of these measurands. The coefficients of this dependence are of the order of 1 pm/με and 10 pm/ºC, respectively. FBG1

FBG2

FBG3

FBG4

FBG5

light source spectrum analysis

λ-range ε

λB1

reflected spectra T

λB2

λB3

λB4

λB5

λ

Fig. 3 Wavelength-division multiplexing of FBG sensors

Therefore, from the measurement of the reflection (or transmission) spectrum of a FBG it is possible to measure strain and temperature. This is a robust absolute measurement sensor that does not have the problems with variations in optical attenuation due to transmission or bending losses in the fiber that plague other OFS types. Moreover, having a spectral measurement system means that it is very easy to use WDM: several FBG with different central wavelengths are deployed on the same fiber to simultaneously measure temperature or strain at several locations by measuring the combined reflection spectrum. This concept is schematically depicted in Fig. 3.

340

A. Loayssa INTERROGATOR

FBG1

Broadband source

Tunable Fabry-Perot

FBG2

FBG4

FBG3

FBG5

FBG6

FBG7

FBG8

FBG10

FBG9

FBG7

FBG8

FBG10

FBG9

Detector

INTERROGATOR

FBG1

Broadband source

FBG2

FBG4

FBG3

FBG5

FBG6

CCD

Polychromator

Fig. 4 Experimental setup for FBG interrogators based on (a) tunable filter and (b) polychromators

Multiple methods for the spectral interrogation of wavelength-multiplexed FBG arrays have been proposed in the literature. However, the most widely deployed commercially are based on tunable filters and lasers, and polychromators. Fig. 4 illustrates the basic experimental setups of these systems. Tunable filter schemes rely on the use of a broadband optical source to illuminate an FBG array [4]. The reflected spectrum is analyzed by using a narrow-band scanning tunable filter, e.g., Fabry-Perot, whose output is fed to a detector. An alternative is to use a fast sweeping laser as a light source instead of a broadband source and avoid the use of the tunable filter. This results in an increase of the measurement dynamic range owing to the larger instantaneous power spectral density reaching the detector as well as an enhancement in the spectral range due to the wider wavelength coverage of the tunable laser. In both cases, an electrical signal is detected as a function of time that can be translated to an optical power versus wavelength trace. Polychromator-based interrogators deploy a very similar setup, but by comparison they measure simultaneously the full reflection spectrum of one or more FBGs by using a diffraction grating to spatially project the wavelength components of the incoming light onto a CCD image sensor. In principle interrogators based on tunable filters or lasers tend to have higher wavelength resolution, and hence sensing resolution, than those based on polychromators. However, the latter offer distinct advantages when it comes to performing dynamic measurements because all components in the reflected spectra are measured at the same time.


341

Commercial interrogators using the techniques just described display wavelength resolution in the pm range, which translates to strain and temperature resolution of the scale of 1 με and 0.1 ºC, respectively. Moreover, dynamic measurements at kHz frequencies are possible. The maximum number of gratings that these systems can interrogate is determined by the total wavelength range of the interrogator divided by the maximum Bragg wavelength shift that the gratings are to accommodate. For instance, using an state-of-the-art interrogator with 80nm wavelength range and using FBG strain sensors with a strain limit of ±2500με and a strain sensitivity around 1pm/με means that a total of 16 grating sensors can be deployed on the same fiber array. In order to increase the total number of measured FBG, interrogators also provide spatial-division multiplexing (SDM) capabilities in which an optical switch is used to select among several sensor channels; hence multiplying the number of sensors by the number of channels, but also increasing the total measurement time. Another possibility is to combine WDM with TDM as depicted in Fig. 5. In TDM the optical source is pulsed and so that time-of-flight information from the reflection of this pulse from multiple FBGs located at increasing distance can be differentiated even if the gratings wavelengths are partially overlapped. The drawbacks of this method are increased measurement time, which reduce its applicability when dynamic measurements are required, and the need to use gratings with low reflectivity (20kΩcm - there is a low likelihood of significant corrosion Between 10 and 20 kΩcm - there is a low to moderate likelihood of corrosion Between 5 and 10 kΩcm - there is a high likelihood of corrosion

New Developments in Sensing Technology for Structural Health Monitoring

New Developments and Applications in Sensing Technology

Structural Health Monitoring

Structural Sensing, Health Monitoring, and Performance Evaluation (Series in Sensors)

Fibre Optic Methods for Structural Health Monitoring

Computational Techniques for Structural Health Monitoring

Fibre Optic Methods for Structural Health Monitoring

Encyclopedia of Structural Health Monitoring

New Trends in Vibration Based Structural Health Monitoring

New developments in parsing technology

New Developments in Parsing Technology

Structural Health Monitoring Using Genetic Fuzzy Systems

Structural Monitoring with Fiber Optic Technology

Joint replacement technology, New Developments

New Developments and Applications in Sensing Technology (Lecture Notes in Electrical Engineering)

New Developments and New Applications in Animal Cell Technology

New Nanotechnology Developments (Nanotechnology Science and Technology)

Rotating Machinery, Structural Health Monitoring, Shock and Vibration, Volume 5

Structural Health Monitoring: with Piezoelectric Wafer Active Sensors

Structural Health Monitoring with Piezoelectric Wafer Active Sensors

Health Monitoring of Structural Materials and Components: Methods with Applications

Health Monitoring of Bridges

Structural Analysis (Mathematics Research Developments)

Artificial Intelligence in Structural Engineering: Information Technology for Design, Collaboration, Maintenance, and Monitoring

New Developments in Differential Equations

New Developments in Biomedical Engineering

New Developments in Advanced Welding

New Developments in Catalysis Research

New Developments in Differential Equations

New Developments in Liquid Crystals

New Developments in Asthma Research

New Developments in Sensing Technology for Structural Health Monitoring

New Developments and Applications in Sensing Technology

Structural Health Monitoring

Structural Sensing, Health Monitoring, and Performance Evaluation (Series in Sensors)

Fibre Optic Methods for Structural Health Monitoring

Computational Techniques for Structural Health Monitoring

Fibre Optic Methods for Structural Health Monitoring

Encyclopedia of Structural Health Monitoring

New Trends in Vibration Based Structural Health Monitoring

New developments in parsing technology

New Developments in Parsing Technology

Structural Health Monitoring Using Genetic Fuzzy Systems

Structural Monitoring with Fiber Optic Technology

Joint replacement technology, New Developments

New Developments and Applications in Sensing Technology (Lecture Notes in Electrical Engineering)

New Developments and New Applications in Animal Cell Technology

New Nanotechnology Developments (Nanotechnology Science and Technology)

Rotating Machinery, Structural Health Monitoring, Shock and Vibration, Volume 5

Structural Health Monitoring: with Piezoelectric Wafer Active Sensors

Structural Health Monitoring with Piezoelectric Wafer Active Sensors

Health Monitoring of Structural Materials and Components: Methods with Applications

Health Monitoring of Bridges

Structural Analysis (Mathematics Research Developments)

Artificial Intelligence in Structural Engineering: Information Technology for Design, Collaboration, Maintenance, and Monitoring

New Developments in Differential Equations

New Developments in Biomedical Engineering

New Developments in Advanced Welding

New Developments in Catalysis Research

New Developments in Differential Equations

New Developments in Liquid Crystals

New Developments in Asthma Research

Recommend Documents